<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by JM Bonthous on Medium]]></title>
        <description><![CDATA[Stories by JM Bonthous on Medium]]></description>
        <link>https://medium.com/@jmbonthous?source=rss-9897f835057e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*SQ_rAmtVWvsk8ouyZlwQFQ.jpeg</url>
            <title>Stories by JM Bonthous on Medium</title>
            <link>https://medium.com/@jmbonthous?source=rss-9897f835057e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 22:34:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jmbonthous/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Hybrid Artist: Reinventing Craft in an Era of Intelligent Tools]]></title>
            <link>https://medium.com/@jmbonthous/the-hybrid-artist-reinventing-craft-in-an-era-of-intelligent-tools-53e91363ca96?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/53e91363ca96</guid>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-artist]]></category>
            <category><![CDATA[ai-art-gallery]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 15:11:01 GMT</pubDate>
            <atom:updated>2026-11-16T15:11:01.658Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>A closer look at the practices and possibilities emerging when creators blend traditional methods with generative technologies.</em></strong></p><p>There’s a false choice floating around every conversation about AI art: you’re either a “traditional artist” or an “AI artist,” as if these are two separate tribes that don’t overlap. Pick a side.</p><p>In practice, almost nobody serious about the medium fits that box. The most interesting work happening right now comes from artists who move fluidly between prompting, painting, coding, and hand-editing, often within a single piece.</p><p>Call this figure the hybrid artist. Part director, part painter, part editor, part technologist — someone who treats AI as one tool among several rather than a replacement for the rest of the toolkit.</p><p>This isn’t a temporary phase on the way to “real” AI art or a retreat back to “real” traditional art. It’s a genuinely new kind of craft, and it’s worth understanding what it actually looks like in practice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*n2-y_3aYGeBZZEF3rDKF1Q.png" /></figure><h3><strong>Prompting as a New Layer of Craftsmanship</strong></h3><p>Dismissing prompting as “just typing a sentence” badly undersells what serious practitioners actually do. A working prompt system is closer to a technical vocabulary than a casual request — specific lighting language, film stock references, compositional terms, built and refined over dozens of iterations.</p><p>Photographers went through exactly this kind of vocabulary-building a century ago. Aperture, exposure, film stock, and developing time all started as unfamiliar technical concepts and became second-nature craft knowledge that photographers stopped consciously thinking about.</p><p>In my own documentary-style AI photography practice, the prompt system didn’t arrive fully formed. It grew across multiple versioned manuals — a working lexicon of terms that reliably produce a specific humanist, black-and-white sensibility, refined the same way a photographer refines their darkroom technique over years.</p><p>Used seriously, prompting behaves less like filling out a form and more like directing a cinematographer: you’re not describing every pixel, you’re communicating intent to a collaborator who has strong instincts of their own and needs precise direction to execute your vision rather than theirs.</p><h3><strong>Iterative Play: Using AI to Explore Ideas at Scale</strong></h3><p>Where AI genuinely earns its place in a hybrid practice isn’t finishing work — it’s exploring ideas faster than any previous tool allowed. A painter sketches a handful of thumbnails before committing to a canvas. A filmmaker storyboards a scene a dozen different ways before shooting.</p><p>AI lets an artist do the equivalent of a hundred thumbnails in the time it used to take to make three. Generate a batch, notice an unexpected mood or composition emerging, and follow that thread rather than the one you originally planned.</p><p>This is closer to brainstorming than production. The value isn’t in any single generated image — it’s in the discovery that happens across the batch, the compositional idea or color relationship you wouldn’t have thought to try on your own.</p><p>Many hybrid artists take exactly that discovery back into traditional media afterward: a composition found through AI exploration becomes the basis for a painting, a physical print, or a hand-built set for photography. The AI proposed a direction; the artist still decides which direction is worth developing.</p><h3><strong>Human Refinement: Where Handwork Still Matters Most</strong></h3><p>Here’s the part that surprises people who assume AI art means the human steps aside: the finishing stage has moved back toward the hand, not away from it. The generated image is increasingly treated as a starting point rather than a final product.</p><p>Concrete techniques are doing real work here — hand-painting texture into an AI base image using Procreate, physically printing a generated image and reworking it with traditional media like ink or gouache, manual color grading and retouching that pushes a digital image toward an analog feel.</p><p>This isn’t a new problem for photography, actually — it’s an old one wearing new clothes. Darkroom photographers have always hand-finished their prints: dodging and burning specific areas, toning the print, adjusting contrast by hand long after the shutter closed. AI just moves the starting point of the process earlier. It doesn’t remove the need for a finishing hand.</p><p>For a lot of hybrid artists, this handwork stage is exactly where their individual voice actually lives. Two artists can generate similar base images from similar prompts and end up with completely different final pieces once their own hand enters the process.</p><h3><strong>Interactive Installations and Living, Generative Works</strong></h3><p>Everything so far has been about a single, eventually-finished image. But some of the most genuinely new hybrid work doesn’t finish at all — installations that generate or respond in real time, shifting based on a viewer’s presence, ambient sound, or live data feeds.</p><p>These works treat the AI model less like a tool used once and more like a live collaborator present in the room, continuously producing variations rather than settling into one final state. A viewer standing in front of the piece today sees something different than a viewer standing there tomorrow.</p><p>This creates genuinely new problems nobody had to solve before. How do you price a work that’s never the same twice? How do you preserve or exhibit something whose content depends on conditions in the room at a given moment, rather than on a fixed object?</p><p>This is the frontier where “hybrid” stops meaning “AI-generated base plus hand-finishing” and starts meaning something structurally different — a collaboration that never fully resolves into a static artifact, closer to a living performance than a painting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*sGn2OdIFNtHz9ToaPezosQ.png" /></figure><h3><strong>How Lower Technical Barriers Lead to Broader Participation</strong></h3><p>Generative tools strip away years of technical training that used to gatekeep visual art entirely — painting fundamentals, darkroom chemistry, years of drawing practice just to render a convincing hand or face. That barrier used to filter who even attempted serious visual work.</p><p>Removing that barrier doesn’t eliminate craft. It relocates where the craft is required — away from technical execution and toward conceptual direction, curation, and taste. The skill shifts from “can you render this” to “can you recognize what’s worth keeping and refining out of everything the tool can produce.”</p><p>This relocation broadens who gets to build a serious practice. Someone without formal art training or expensive equipment can now develop a genuinely distinctive visual language, provided they put in the less visible work of developing judgment rather than manual technique.</p><p>Lower barriers to entry get mistaken constantly for lower standards. In practice, the standard hasn’t dropped — it’s simply moved to a stage of the process that’s much less visible from the outside than a paintbrush or a camera.</p><p><strong>Defining Artistic Value When Creation Is Shared With a System</strong></p><p>All of this circles back to one unresolved question: if a model contributes genuine, unpredictable creative input, how do we assign value and authorship to work made jointly with it?</p><p>The clearest working answer sitting in front of us: value belongs to the artist who builds a consistent, recognizable practice across the entire hybrid process — the prompting vocabulary, the exploratory batches, the hand-finishing choices, the sequencing of a body of work — not to any single output the system happened to produce.</p><p>That’s a familiar answer, actually. It’s the same one art history has always given for collaborative work, just applied to a new kind of collaborator. The value was never really in the pigment or the shutter click. It was always in the eye directing the whole process.</p><p>The hybrid artist isn’t a transitional figure standing between the old world and some future, fully-automated one. Part director, part painter, part editor — this blended practice may simply be what “artist” comes to mean from here forward.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=53e91363ca96" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seeds, Weights, and Negative Prompts: Controlling Variation Instead of Gambling]]></title>
            <link>https://medium.com/@jmbonthous/seeds-weights-and-negative-prompts-controlling-variation-instead-of-gambling-303154a58f29?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/303154a58f29</guid>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-art-gallery]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai-artist]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 15:06:01 GMT</pubDate>
            <atom:updated>2026-11-09T15:06:01.495Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>How to iterate on purpose instead of rerolling until something works</em></strong></p><p>At two in the morning, Bram Takahashi-Hale got the image he’d been chasing all week: a lighthouse in a storm, the beam cutting through rain at exactly the right angle.</p><p>He wanted the sky a little darker, so he changed four words and generated again. The lighthouse moved. He changed them back and generated again. A different lighthouse. By 2:15 the image was gone for good.</p><p>Ondine Marchetti, a printmaker who shares his studio, watched the whole thing. To her, it looked like pulling a perfect proof and then wiping the plate clean before writing down the ink mix.</p><p>Bram was gambling. He changed the prompt and the randomness at the same time, every time, so he could never tell what caused what. Every good image was luck, and luck doesn’t repeat.</p><p>Seeds, weights, and negative prompts are the controls that turn rerolling into a method. The principle is the one Ondine uses for every edition: lock everything, then change one thing at a time.</p><h3>Why Rerolling Feels Productive but Isn’t</h3><p>Every image generation starts from random noise, a field of static that the model gradually shapes into a picture. That’s why the same prompt gives you a different image each time you press the button.</p><p>Randomness is wonderful for exploring. It’s terrible for refining. When you change your prompt and get new randomness at the same time, two variables move at once, and you can’t know which one made the image better or worse.</p><p>The result is a familiar pattern: lucky accidents you can’t repeat, and very little learning from one session to the next. Hundreds of generations, and the skill doesn’t grow.</p><p>Ondine’s analogy is exact. No printmaker changes the ink, the paper, and the press pressure all at once and expects to know which one fixed the proof. You change one, pull a proof, and compare.</p><h3>Seeds: Locking the Starting Point</h3><p>A seed is the number that sets that starting noise. Use the same seed, with the same prompt, settings, and model, and you get the same image, or something very close to it.</p><p>Where you find it depends on the tool. Stable Diffusion interfaces show the seed for every image. Midjourney lets you set one with a seed parameter. Some hosted tools don’t expose seeds at all, which tells you how much control they’re offering.</p><p>One caveat: not every tool reproduces an image exactly from the same seed. Some only give you something similar. Test yours once so you know what to expect.</p><p>Here’s how to use it. Explore with random seeds first, and when you find a composition you like, note its seed and lock it. From then on, every change you make to the prompt is a controlled experiment, because the starting point stays the same.</p><p>A seed isn’t a style, though. It carries composition well only while the prompt stays close to the original. Rewrite half the prompt, and the seed matters much less.</p><p>Ondine labels every keeper with its seed, the way she numbers an edition. It takes two seconds, and it’s the difference between a lucky image and one you can return to.</p><h3>Weights: Telling the Model What Matters Most</h3><p>Not every word in a prompt carries equal weight, and you can shift the balance. In Midjourney, multi-prompts separate concepts with a double colon and give each a number: “lighthouse::2 storm::1” makes the lighthouse twice as important as the storm.</p><p>Stable Diffusion interfaces use a different syntax: a word or phrase in parentheses with a value, such as “(dark storm clouds:1.3).” Values above 1 strengthen a term, and values below 1 soften it.</p><p>Some newer models ignore weighting syntax entirely. With those, emphasis comes from plain language: put the important element first, describe it more clearly, or mention it again. Check how your model behaves before relying on numbers.</p><p>Reference weights work the same way for images. Image weight, style reference weight, and character reference weight decide how strongly an uploaded picture controls the result. Raise them when the reference is being ignored, and lower them when it’s being copied.</p><p>Guidance, often called CFG, controls how literally the model follows your prompt overall. Too low, and the image wanders off-brief. Too high, and it looks overcooked, with harsh contrast and oversaturated color.</p><p>The practical rule is the same for all of these: adjust one weight at a time, in small steps, on a locked seed.</p><h3>Negative Prompts: Removing Without Summoning</h3><p>A negative prompt tells the model what to steer away from. Stable Diffusion tools give you a separate field for it. Midjourney offers a “no” parameter, which works like giving that word a negative weight.</p><p>The classic trap is putting the negative in the main prompt. Write “a beach, no people,” and you’ll often get people, because the model registers the word “people” far more strongly than the word “no.”</p><p>The second trap is the giant boilerplate list copied from a forum: forty terms like “ugly, blurry, deformed, extra fingers.” Long lists tend to dull the image and can fight your main prompt. Use a few targeted terms for problems you actually see.</p><p>Many models handle negatives poorly, or don’t support them at all. With those, describe what you want instead of what you don’t: “an empty beach at dawn, footprints washed away” rather than “a beach, no people.”</p><p>Ondine’s rule is a good one. She adds a negative only after she’s seen the same defect three times. Anything less is guessing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/1*4ddvv-lcVKGgk8c4BdZIIw.png" /></figure><h3>The Other Variation Dials</h3><p>Most tools also offer variation controls, often labeled subtle and strong. Subtle variations make small moves around an image you already like: the same composition with slightly different details. Strong variations wander further.</p><p>Midjourney adds two more dials. Chaos controls how different the images in a batch are from one another, and stylize controls how much of the tool’s own aesthetic it layers on top of your prompt.</p><p>Remix and edit modes let you change the prompt while keeping a chosen image’s structure. They’re useful when the composition is right but the season, light, or medium is wrong.</p><p>Think of these in two groups. High chaos, strong variations, and random seeds are exploration dials. Subtle variations, a locked seed, and remix are refinement dials. Bram had been using exploration dials for refinement work, which is why nothing ever settled.</p><h3>A Five-Step Iteration Method</h3><p>Put together, these controls make a repeatable method. Step one is to explore. Use random seeds, high variety, and loose prompts. Generate widely, then select ruthlessly. This is the only stage where rerolling is the right move.</p><p>Step two is to lock. Choose one result, record its seed and settings, and freeze them. From here on, the starting point doesn’t move.</p><p>Step three is to refine. Change one thing per round, whether a word, a weight, or a single negative, and compare the result with the previous one side by side. If it’s better, keep it. If not, undo it.</p><p>Step four is to vary. Once the image is close, use subtle variations or nearby seeds to generate alternatives around the winner. You’re choosing between siblings, not strangers.</p><p>Step five is to record. Save the prompt, seed, settings, and model version alongside the final file, so you can return to the image, extend it, or build a series from it.</p><p>Bram tried it the following week. He made thirty generations instead of three hundred and ended up with a lighthouse he could darken, re-light, and turn into a set of four.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QQevXqbjgZSLFuZ5e4OIpg.png" /></figure><h3>Keep an Iteration Log</h3><p>The last habit is the simplest. Keep a log: a spreadsheet or plain text file with the image name, seed, prompt, settings, model version, and a one-line note on what you changed and why.</p><p>Ondine calls it her proof record, the same thing she keeps for every edition she prints. It lets her reproduce a result months later, extend a single image into a series, and see her own patterns over time.</p><p>It pays off in another way, too. A record of your decisions is solid evidence of authorship if a gallery, client, or platform ever asks how a piece was made.</p><p>Bram still loves happy accidents. The difference now is that when one arrives, he can catch it, keep it, and build on it.</p><p>That’s the whole method: explore freely, then lock everything and change one thing at a time. Iterate on purpose, and every image teaches you something.</p><p><em>Jean-Marie Bonthous (publishing as J.M. Bonthous) is the author or more than two dozen books, including six on message-driven marketing, seven on the human dimensions on AI and several on AI art and AI filmmaking</em>. <em>He writes several blogs on Medium — About the human dimensions of AI:</em> <a href="https://medium.com/@jm_26203"><em>AI in Real Life</em></a>,<em> About AI art:</em> <a href="https://medium.com/@jmbonthous"><em>The Algorithmic Eye</em></a>,<em> About AI filmmaking:</em> <a href="https://medium.com/@info_13818"><em>The Solitary Frame</em></a><em>, About AI Marketing: </em><a href="https://medium.com/@jm_b_73712"><em>The AI Marketer.</em></a><em> See his latest books at </em><a href="http://www.jmbonthous.com/"><em>www.jmbonthous.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=303154a58f29" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Value in the Age of Infinite Images: How AI Art Disrupts the Market]]></title>
            <link>https://medium.com/@jmbonthous/value-in-the-age-of-infinite-images-how-ai-art-disrupts-the-market-1097cb665230?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/1097cb665230</guid>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[ai-art-gallery]]></category>
            <category><![CDATA[ai-artist]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 15:06:01 GMT</pubDate>
            <atom:updated>2026-11-02T15:06:01.652Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>The economics of creativity are being rewritten as generative tools shift the balance between scarcity, labor, and collectibility.</em></strong></p><p>Art has always been priced on scarcity. A painting takes weeks to finish, so there’s only one of it. A photographer prints a limited edition of twenty, then destroys the negative or retires the file. Time and rarity have always been baked into the number on the price tag.</p><p>Now imagine an artist who can generate a thousand finished images in a single afternoon, at a cost of pennies each. That’s not a hypothetical — it’s a Tuesday for anyone working seriously with AI art tools today.</p><p>The old pricing logic doesn’t know what to do with that. If scarcity and labor no longer explain value, something else has to. Nobody in the market has fully worked out what that something else is yet.</p><h3><strong>From Labor-Intensive Craft to Near-Instant Output</strong></h3><p>Traditional art pricing has always been, at its core, a labor calculation dressed up in aesthetic language. A gallery pricing an oil painting is implicitly pricing weeks of the artist’s time, the cost of materials, and the fact that this exact canvas exists nowhere else.</p><p>A darkroom photographer selling a print is selling something similarly scarce — a physical object, hand-processed, from an edition capped at a small number. Ansel Adams himself printed fewer copies of his most famous images than most people assume, and that scarcity is part of why they command the prices they do.</p><p>AI generation removes almost the entire labor input from that equation. The heavy editing and curation work still exists, as any serious AI artist knows, but the raw generation step — the part that used to take hours of darkroom time or weeks of brushwork — now takes seconds.</p><p>This creates an uncomfortable question for anyone trying to price AI work using the old formula: if an image took ten seconds to generate, what exactly is a collector paying for?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/978/1*Z3WUaJYJ4qs9knaEj0aJnw.png" /></figure><h3><strong>The Price Problem: Can AI Art Hold Market Value?</strong></h3><p>The market has already run this experiment once. In 2018, Christie’s sold an AI-generated portrait, “Edmond de Belamy,” for $432,500 — a price that stunned the art world and made headlines everywhere.</p><p>What followed was less triumphant. Later AI works by the same collective sold for far less, and critics increasingly described the original sale as driven by novelty rather than lasting artistic merit. The shock value of “a machine made this” faded fast once AI-generated images became commonplace.</p><p>That’s the durability problem in a nutshell. A rare hand-made object tends to hold or increase in value over decades. A novelty tends to depreciate the moment the novelty wears off — and generative tools have gone from headline-grabbing curiosity to something a teenager does on their phone in under ten years.</p><p>If value can’t come from novelty and can’t come from labor scarcity, the more promising place to look is process and concept — not the image itself, but the thinking, curation, and distinctive voice behind it. That shift is already visible in how serious AI artists are choosing to sell their work.</p><h3><strong>Collectors Confront a Flood of Machine-Generated Work</strong></h3><p>Put yourself in a collector’s position. You used to be able to judge a piece partly on visible technical mastery — brushwork, printing quality, evidence of years of practice. Those signals told you something about scarcity and effort even before you knew anything about the artist.</p><p>AI removes most of those signals. A technically flawless image tells you almost nothing about how much thought went into it, since the model can produce technical polish on the first try. Skill and effort are no longer visible on the surface of the image the way they used to be.</p><p>New signals are emerging to fill that gap. Reputation and a documented, consistent artistic voice matter more than ever — collectors increasingly look for an artist who has clearly built a specific visual world over dozens or hundreds of works, not someone who generated one striking image and stopped.</p><p>Some artists are deliberately reintroducing scarcity by hand: capping editions at a fixed number, numbering physical prints, writing artist statements that document the prompt system and editing process behind a piece. These are the same tools photography used a century ago to make an infinitely reproducible medium feel collectible — and they’re being redeployed for the same reason now.</p><h3><strong>Institutions Struggle to Judge Significance Amid Abundance</strong></h3><p>Museums and galleries have always functioned as filters. Limited wall space and a handful of acquisitions per year forced curators to make hard choices, and those choices became a kind of quality signal collectors trusted.</p><p>Abundance breaks that filtering mechanism at the source. With millions of AI images produced every day across every platform, the old model of a curator patiently reviewing a manageable stream of submissions simply doesn’t scale to the volume involved.</p><p>Many institutions have responded with caution rather than adaptation — moving slowly on acquiring AI work, partly because of unresolved questions about training data and copyright, and partly because nobody wants to be the museum that bought a piece history later dismisses as a gimmick.</p><p>The result is a widening gap between where the actual market activity is happening — artist-direct sales, online marketplaces, niche digital-art platforms — and where the traditional validation of “this matters” still lives. Serious AI artists are increasingly building reputations outside the gatekeeping structures that used to define art-world legitimacy.</p><h3><strong>The Uneasy Coexistence of Human and Machine Creativity</strong></h3><p>None of this means AI art is replacing the traditional market. What’s actually happening looks more like two markets developing side by side, occasionally overlapping, often eyeing each other with suspicion.</p><p>Traditional collectors and dealers remain broadly skeptical, still anchored to scarcity-and-labor pricing logic that AI genuinely doesn’t fit. A newer, often younger and more digitally native buyer base treats AI art as its own category, with its own value logic borrowed partly from digital collectibles culture.</p><p>The most interesting bridge between the two worlds is the artist who treats generation as only the first stage of a much longer, still labor-intensive process — extensive post-generation editing, hand-finishing, careful sequencing into bodies of work. That kind of practice reintroduces exactly the time investment and distinctive voice that traditional collectors know how to value.</p><p>The real friction isn’t “human creativity versus machine creativity.” It’s an old value system built on scarcity colliding with a new reality built on abundance, and both sides are still negotiating the terms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4mEHvUwEPIpFBTXFNRDJIw.png" /></figure><h3><strong>What the Future of Art Buying Might Look Like</strong></h3><p>The most likely direction: value keeps migrating toward whatever can’t be copied at zero cost. A model can be prompted by anyone. An artist’s specific, hard-won visual voice — developed over years and hundreds of works — cannot.</p><p>Expect new collectibility markers to matter more: documented, verifiable process (which system, which methodology, which edition), authenticated limited runs, and a “canon” body of work rather than any single standout image. Collectors will increasingly buy into a practice, not just a picture.</p><p>The market will likely split into two clear tiers: a high-volume, low-value flood of casual AI output that behaves like a commodity, and a smaller, curated tier of serious practitioners whose work behaves much more like traditional fine art — priced on reputation, process, and physical presence rather than novelty.</p><p>The flood of images doesn’t destroy value. It relocates it — away from the object itself, and onto the artist standing behind it.</p><p><em>Jean-Marie Bonthous (publishing as J.M. Bonthous) is the author or more than two dozen books, including six on message-driven marketing, seven on the human dimensions on AI and several on AI art and AI filmmaking</em>. <em>He writes several blogs on Medium — About the human dimensions of AI:</em> <a href="https://medium.com/@jm_26203"><em>AI in Real Life</em></a>,<em> About AI art:</em> <a href="https://medium.com/@jmbonthous"><em>The Algorithmic Eye</em></a>,<em> About AI filmmaking:</em> <a href="https://medium.com/@info_13818"><em>The Solitary Frame</em></a><em>, About AI Marketing: </em><a href="https://medium.com/@jm_b_73712"><em>The AI Marketer.</em></a><em> See his latest books at </em><a href="http://www.jmbonthous.com/"><em>www.jmbonthous.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1097cb665230" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inpainting and Outpainting: A Painter’s Guide to Fixing One Area Without Wrecking the Rest]]></title>
            <link>https://medium.com/@jmbonthous/inpainting-and-outpainting-a-painters-guide-to-fixing-one-area-without-wrecking-the-rest-36b98e40ef95?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/36b98e40ef95</guid>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[digital-art]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-image-generator]]></category>
            <category><![CDATA[ai-art]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 15:11:01 GMT</pubDate>
            <atom:updated>2026-10-26T15:11:01.898Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Masking, blending, and matching light and texture, so the repair disappears</em></strong></p><p>Benedikt Saar had a nearly perfect portrait: a woman at a window, soft afternoon light, a quiet expression. And six fingers on her left hand.</p><p>He masked the hand and regenerated it, and the new hand was flawless: five fingers, elegant knuckles, crisp detail. It also looked like a sticker. It was sharper than the face, cooler than the skin around it, and lit from the wrong side.</p><p>Odalys Ferreira-Kahn, an oil painter who had started working with AI a few months earlier, saw it from across the room. She didn’t know what a denoising slider was, but she knew exactly what had gone wrong, because she’d made the same mistake on canvas years ago.</p><p>Repaint one passage on a dry painting without glazing it back into the whole, and it jumps forward. Every painter learns this the hard way. Every passage has to agree with its neighbors in light, edge, texture, and temperature.</p><p>That’s the whole secret of inpainting. The tool is new; the principle is centuries old. Fix the part, but judge the whole. Here’s how to do it, with Odalys’s eye and Benedikt’s settings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uXSn28dxBUbcw8bjo69c9A.png" /></figure><h3>What These Tools Do</h3><p>Inpainting regenerates a masked area of an image. You paint over the part you want changed, describe what should be there, and the model fills it in, using the rest of the image as its guide.</p><p>Outpainting extends the canvas beyond its original edges. Same idea, but the model invents new territory instead of replacing old territory: more sky, more room, a wider frame.</p><p>You’ll find these tools almost everywhere now: Generative Fill in Photoshop, region editing in Midjourney, instruction-based editors like Flux and Gemini, and full manual control in Stable Diffusion interfaces such as ComfyUI. Names and features change often, so check what your version offers.</p><p>The key idea in all of them is context. The model only sees what surrounds your mask. Light, color, texture, and perspective are all inferred from the neighboring pixels.</p><p>That means your mask is not just a selection. It’s a decision about what the model gets to look at. Most bad repairs start right there.</p><h3>Why Repairs Look Pasted On</h3><p>The sticker look almost always comes from one of four mismatches. The first is light: the new area is lit from a different direction, has a different color temperature, or casts shadows that are too hard or too soft.</p><p>The second is texture. Generated patches tend to be cleaner and smoother than what surrounds them. In a painterly image, the brushwork suddenly vanishes; in a photographic one, the grain does.</p><p>The third is resolution. The patch is often generated at a different effective sharpness, so it reads as crisper or blurrier than the rest of the picture, even when the content is right.</p><p>The fourth is the edge. A hard mask produces a hard seam, a faint line where the old image ends and the new one begins, and the eye finds it instantly.</p><p>Odalys recognized all four on Benedikt’s hand. In painting terms, the passage was repainted on a dry canvas with the wrong light, fresh paint, and a hard edge. The fixes are the same in both worlds.</p><h3>Mask Like a Painter, Not a Surgeon</h3><p>Benedikt’s instinct was surgical: mask the six fingers tightly and nothing else. Odalys’s instinct was the opposite. She’d mask the hand, the shadow it cast on the sill, and a little of the sleeve, because those things are all part of the same passage.</p><p>That’s the first rule. Mask slightly beyond the defect, and include whatever the object touches or affects: its shadow, its reflection, the edge where it meets the background. If those stay frozen, the new object won’t fit them.</p><p>Second, soften the mask edge. Most tools offer feathering or mask blur; use it. A soft boundary lets the model blend the old and new pixels instead of stitching them together.</p><p>Third, think in lost and found edges, the painter’s term for edges that dissolve versus edges that stay sharp. End your mask inside a soft area, like a shadow or an out-of-focus background, never exactly on a hard edge like a jawline or a window frame.</p><p>Fourth, fix one problem per pass. A large mask invites the model to reinvent things you liked. Small, targeted masks keep the rest of the image yours.</p><p>If your tool lets you choose how much surrounding image the model can see, sometimes labeled padding or “whole picture” versus “only masked,” give it enough context to read the light. A hand regenerated without seeing the window will never be lit by it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QUBJ9FYL0HXAOSdXZUn-vg.png" /></figure><h3>The Settings That Matter</h3><p>The most important control is strength, called denoising strength in many tools. It decides how much of the original the model keeps. Low values refine what’s there; high values throw it away and start fresh.</p><p>As a starting point, try something around 0.3 to 0.5 when you want to correct an area while keeping its shape and color, and 0.7 or higher when you need to replace it entirely. Every tool scales this differently, so treat these numbers as a range to test, not a rule.</p><p>Benedikt’s habit was maximum strength every time. That’s exactly why his hand looked pasted on: the model ignored the original light and color and invented its own.</p><p>Prompt the patch, not the picture. Describe only what belongs inside the mask, then repeat the image’s light and medium in plain words: “soft window light from the left, warm skin, visible oil brushwork.” Those few words do more for blending than any slider.</p><p>Finally, generate several variations and choose. Selection beats re-prompting. Four or six options usually include one that already agrees with its surroundings.</p><h3>Match Light, Color, and Texture</h3><p>Before you inpaint, read the light the way a painter would. Where is it coming from? Is it warm or cool? Hard, with crisp shadows, or soft, with gentle falloff? Put those answers in your patch prompt.</p><p>After you inpaint, check the values. Squint at the image, or convert a copy to grayscale. The repaired area should sit in the same range of lights and darks as its neighbors. If it’s brighter or darker, the eye will find it no matter how good the color is.</p><p>Fix color with a light hand. Sample the surrounding area, then apply a curves or color-balance adjustment limited to the patch. Odalys borrowed a term from her studio: glaze it. A low-opacity color layer over the repair pulls it back into the whole, just as a thin transparent glaze does on canvas.</p><p>Then match the texture. Generated patches are usually too smooth, so add a little noise or grain to the patch alone. For painterly work, overlay a brushstroke or canvas texture, or paint over the seam by hand, digitally or on the print.</p><p>Last, match sharpness. Soften or sharpen the patch to agree with its surroundings, not to look “better.” A repair that’s sharper than the face around it is still a sticker.</p><h3>Outpainting Without Seams</h3><p>Outpainting fails in its own ways, and the biggest mistake is extending too far at once. Grow the canvas in steps, roughly a quarter to a half of the image at a time, and let each new section overlap the existing image so the model has something to continue.</p><p>Keep the geometry consistent. The horizon line, vanishing points, and light source must carry across the extension. Painters square up a composition with guide lines before enlarging it; you can do the same on a separate layer.</p><p>Watch the new edges for repetition. Models like to clone what they already see, so a new stretch of shoreline may quietly repeat the same three rocks. Small, targeted inpaints fix that.</p><p>Then recompose. Outpainting changes the composition, sometimes dramatically. Benedikt widened a portrait into a landscape format and suddenly the woman sat dead center, with the focal point lost in empty wall.</p><p>Odalys’s fix was simple: crop again. Extend generously, then choose the frame. The extension is raw material, not the final composition.</p><h3>The Step-Back Test</h3><p>Painters step back from the easel constantly, because a passage that looks right from six inches can look wrong from six feet. Do the digital version: view the whole image small, then full screen, then printed if it’s headed for a wall.</p><p>Then flip it horizontally. Painters use a mirror for this, because a reversed image looks fresh and the eye catches what it had stopped seeing. Seams, mismatched light, and sticker patches jump out immediately.</p><p>Use a five-second rule. If you can find the repair within five seconds of looking at the whole image, so can a viewer. Go back and glaze, blend, or re-mask.</p><p>By the end of the month, Odalys had learned the sliders and Benedikt had learned to step back. His second attempt at the hand took ten minutes: a wider mask that included the shadow, moderate strength, “soft window light from the left” in the prompt, and a little grain. Nobody could find it.</p><p>The tools will keep changing. The principle won’t. Fix the part, but judge the whole.</p><p><em>Jean-Marie Bonthous (publishing as J.M. Bonthous) is the retired founder of a successful Digital Marketing agency. He is the author or more than two dozen books, including six on message-driven marketing, seven on the human dimensions on AI and several on AI art and AI filmmaking</em>. <em>He writes four blogs on Medium — About the human dimensions of AI:</em> <a href="https://medium.com/@jm_26203"><em>AI in Real Life</em></a><em> About AI art:</em> <a href="https://medium.com/@jmbonthous"><em>The Algorithmic Eye</em></a><em> About AI filmmaking:</em> <a href="https://medium.com/@info_13818"><em>The Solitary Frame</em></a><em> About Message-Driven Marketing: </em><a href="https://medium.com/@jm_b_73712"><em>The Message-Driven Marketer</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=36b98e40ef95" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Your Sketchbook to the Model: Using Your Own Drawings as AI Input]]></title>
            <link>https://medium.com/@jmbonthous/from-your-sketchbook-to-the-model-using-your-own-drawings-as-ai-input-cfa7a2ecffe1?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/cfa7a2ecffe1</guid>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[ai-image-generator]]></category>
            <category><![CDATA[digital-art]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-art]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 15:06:01 GMT</pubDate>
            <atom:updated>2026-10-19T15:06:01.758Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Image prompts, sketch-to-image tools, and how much control you actually keep</em></strong></p><p>Marguerite Oyelaran has filled sketchbooks for twenty years. Last spring she uploaded one of her favorite drawings, a Lisbon tram climbing a steep street, drawn on location in pen and wash.</p><p>The model gave her back a beautiful tram. It wasn’t hers. Her low, tilted viewpoint had flattened into a postcard angle, the street had straightened out, and the crooked lamppost she loved was gone.</p><p>She felt the way many painters do the first time: the machine had looked at her drawing and politely ignored it.</p><p>That’s the misconception at the heart of sketch-to-image work. Uploading a drawing doesn’t mean the model will respect it. In many tools, your image is only a suggestion, unless you use a method that treats it as structure.</p><p>The good news is that you can choose how much control you keep. The skill lies in deciding which parts of your drawing are law and which are suggestions, then picking the tool that enforces that decision.</p><h3>The Control Spectrum</h3><p>Every way of feeding a drawing to a model sits somewhere on a spectrum, from “loosely inspired by” to “built on top of.” Knowing where your tool sits saves hours of frustration.</p><p>The loosest level is the image prompt. The model borrows your drawing’s subject, mood, and palette, and only roughly its composition. Midjourney’s image prompts work this way, with an image weight setting to raise or lower their influence. This level is great for inspiration and poor for preserving a drawing.</p><p>The middle level is image-to-image. The model repaints directly over your drawing, and a strength setting decides how much of the original survives. At low strength it keeps your lines and adds polish; at high strength it keeps little more than a general arrangement.</p><p>The tightest level is structure-guided generation. The model follows the lines, edges, or depth of your drawing while inventing everything else: texture, light, and color. ControlNet modes such as scribble, lineart, and depth work this way in Stable Diffusion tools, as do structure-reference features in Adobe Firefly and real-time sketch canvases like Krea.</p><p>Instruction-based editors such as Flux Kontext and Gemini sit somewhere in between. You can tell them, “render this sketch as a watercolor and keep the composition,” and they often comply, though not always faithfully.</p><p>Feature names and availability change often, so check what your tool currently offers. The spectrum itself doesn’t change.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sWqQ_2DjTxoRwx6ANlOvFg.png" /></figure><h3>Decide What’s Law and What’s Suggestion</h3><p>Before you upload anything, write two short lists. The first is what must survive: the composition, the pose, the perspective, the silhouette, or one specific detail you care about.</p><p>The second is what you’re willing to hand over. Usually that means rendering, texture, lighting, and color, the parts the model does well and that you might enjoy seeing reinterpreted.</p><p>Then match the lists to the spectrum. If composition is law, you need image-to-image at moderate strength or structure guidance. If mood and palette are all you want, an image prompt is fine.</p><p>Marguerite’s mistake was simple: she used the loosest level for the tightest goal. Her viewpoint and her lamppost were law, and she handed them to a tool that treats everything as a suggestion.</p><h3>Prepare Your Sketch</h3><p>Models read every mark as intentional, so prepare the drawing before you upload it. Scan it, or photograph it flat in even light, then remove shadows, the paper’s tint, and any notebook lines. A gray, uneven background often turns into fog or texture.</p><p>Raise the contrast until the lines read clearly. Faint pencil, which looks delicate to a person, often reads as noise to a model. Erase stray marks and smudges you don’t want rendered as objects.</p><p>Crop to the final aspect ratio before you upload, not after. A square sketch forced into a widescreen frame invites the model to invent the missing space, and that’s where compositions drift.</p><p>Match the drawing to the job. A value thumbnail with three or four tones works well for image-to-image. A clean contour drawing suits line-based control. A quick gesture drawing is best when the pose is what matters.</p><p>Lazlo Pereira-Wynn, an AI native who has never trained as an artist, found this out by accident. A crude thirty-second thumbnail of three gray shapes gave him more control over composition than any prompt he’d ever written. Composition is made of shapes, not words.</p><h3>The Settings That Decide How Much Survives</h3><p>In image-to-image, the strength setting is everything. Low values keep your drawing’s actual look, mid values keep its structure while repainting the surface, and high values keep almost nothing. Every tool scales this differently, so run the same sketch at three or four strengths and compare the results.</p><p>In structure-guided tools, the equivalent is control weight, which sets how strictly the model follows your lines. Set it too high and the result looks like a coloring book filled in by a careful child. Set it too low and your drawing drifts away again.</p><p>Many tools also let you apply control only during the early part of generation. That locks in your composition first and then frees the model to render naturally. It’s often the best compromise between faithful and alive.</p><p>Your text prompt still matters. Describe the rendering you want, such as “loose gouache, morning light, warm shadows,” rather than restating what’s already in the drawing. The drawing handles the what; the prompt handles the how.</p><p>Then generate in batches and choose. Six variations at the right settings will nearly always include one that respects your drawing better than the rest.</p><h3>The Drawing Loop</h3><p>Don’t stop at one pass. Print the result or open it in a drawing app, draw your corrections directly on top, and feed the corrected image back in.</p><p>This is where traditional artists gain the most. The work becomes a conversation between your hand and the model rather than a one-way handoff. You redraw what matters, and the model re-renders around it.</p><p>Marguerite’s second attempt took three loops. She used structure guidance to lock her viewpoint, redrew her crooked lamppost by hand on the first output, and ran it again at moderate strength. The final image was unmistakably her tram.</p><p>Some artists go one step further and finish by hand: drawing or painting directly on the final print. It’s the strongest way to make the work visibly yours, and it adds a physical layer no model can reproduce.</p><h3>How Much Is Still Yours?</h3><p>This question matters to both groups of artists, and the practical answer is simple: the more of your own drawing is visibly preserved in the final image, the stronger your claim to it, artistically and potentially legally.</p><p>The U.S. Copyright Office’s 2025 report on AI and copyrightability addressed this directly. It indicated that when an artist’s own expressive work, such as a drawing, remains perceptible in the output, that human contribution can be protected, even if the AI-generated additions on their own are not. Guidance is still evolving and varies by country, so check the current position, or ask a lawyer if a sale depends on it.</p><p>Either way, keep records. Save your original sketch, every intermediate file, the settings you used, and any hand corrections. If your authorship is ever questioned by a gallery, a client, or a platform, a documented process is your best answer.</p><p>The honest scale looks like this: with image prompts alone, you’re mostly a curator. With structure guidance and a drawing loop, you’re mostly an author. Choose the level that matches the work you want to claim.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sScrMjT8c9EvLkmtcV1hsg.png" /></figure><h3>Quick Fixes for Common Failures</h3><p>If the model ignores your composition, move up the spectrum: switch from an image prompt to image-to-image or structure guidance, or raise the control weight.</p><p>If the result looks traced and lifeless, do the opposite. Lower the control weight, or end the control earlier in the generation so the model has room to render naturally.</p><p>If hands and faces come out strange, draw them more clearly in the sketch, or leave them loose and fix them afterward with inpainting. If your paper’s texture or tint shows up in the output, go back and clean the scan.</p><p>Lazlo learned that a crude drawing can be a more powerful prompt than a perfect sentence. Marguerite learned to choose her tool by what she refuses to lose.</p><p>Your sketchbook is the most personal input you’ll ever give a model. Decide what’s law, then pick the tool that enforces it.</p><p><em>Jean-Marie Bonthous (publishing as J.M. Bonthous) is the retired founder of a successful Digital Marketing agency. He is the author or more than two dozen books, including six on message-driven marketing, seven on the human dimensions on AI and several on AI art and AI filmmaking</em>. <em>He writes four blogs on Medium — About the human dimensions of AI:</em> <a href="https://medium.com/@jm_26203"><em>AI in Real Life</em></a><em> About AI art:</em> <a href="https://medium.com/@jmbonthous"><em>The Algorithmic Eye</em></a><em> About AI filmmaking:</em> <a href="https://medium.com/@info_13818"><em>The Solitary Frame</em></a><em> About Message-Driven Marketing: </em><a href="https://medium.com/@jm_b_73712"><em>The Message-Driven Marketer</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cfa7a2ecffe1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Camera and Lens Language That Works in Prompts]]></title>
            <link>https://medium.com/@jmbonthous/camera-and-lens-language-that-works-in-prompts-abc93a1915c7?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/abc93a1915c7</guid>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[ai-artist]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[digital-art]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 14:06:01 GMT</pubDate>
            <atom:updated>2026-10-12T14:06:01.452Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>What “85mm,” “f/1.4,” “tilt-shift,” and “overhead” really change in the image, and what they don’t</em></strong></p><p>Sabine Okonkwo-Reyes spent fifteen years as a press photographer before she started making AI images. Her prompts read like spec sheets: “Canon 5D, 85mm, f/1.4, ISO 100, 1/250s, natural light.”</p><p>Thaddeus Vane is a painter who has never owned a serious camera. His prompt for a similar portrait was one plain sentence: “face sharp, background melting into soft warm blur, seen from slightly below.”</p><p>They ran both prompts side by side. Thaddeus’s portrait looked more like what Sabine had in mind. Half her specs, it turned out, had changed nothing at all.</p><p>The reason is simple. An image model has no lens and no physics. It learned camera words from the captions of millions of real photographs, so a term only works when it has a strong, consistent visual signature in that data.</p><p>The rule that follows: prompt for the visible effect, and use camera language only where the camera term is the shortest route to that effect. Here’s which terms earn their place, which don’t, and how to tell the difference yourself.</p><h3>How Models “Understand” Camera Language</h3><p>When a model sees “85mm,” it doesn’t calculate a field of view. It recalls what images captioned with “85mm” tended to look like: head-and-shoulders portraits, flattering proportions, and softly blurred backgrounds.</p><p>That’s association, not simulation. It explains why some camera terms work beautifully and others do nothing. A term with a strong, common visual look pulls the image in a clear direction. A term with no visible trace in the photo has nothing to pull with.</p><p>ISO is a good example. In a real camera it changes noise and exposure, but in most photos people share, you can’t see which ISO was used. So the model has little to associate with “ISO 100,” and the term barely moves the result.</p><p>For traditional artists, here’s a useful translation: treat each camera term as shorthand for a bundle of visual decisions about space, edges, and focus. You already make those decisions with a brush. Camera language just names them in a way the model has learned.</p><p>With that in mind, let’s go through the terms that matter, starting with the one that shapes space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d6WYfWeDOH6m5uSauKLdsQ.png" /></figure><h3>Focal Length: How Space Feels</h3><p>Wide lenses, roughly 14 to 24mm, exaggerate depth. Foregrounds loom, edges stretch, and the scene recedes dramatically. They’re useful for interiors, sweeping landscapes, and anything that should feel immersive or slightly uneasy.</p><p>At 35mm you get the documentary, street-photography feel: the subject plus its surroundings. At 50mm, the look is close to natural vision, neutral and unremarkable in the best sense.</p><p>From about 85 to 135mm, you’re in classic portrait territory. Faces look flattering, and the background is pulled closer and softened. Beyond 200mm, compression gets strong: distant layers stack flat, which is how photographers make a city skyline seem to sit right under a mountain.</p><p>Painters will recognize the difference. Wide lenses give you deep, receding perspective. Long lenses give you flattened, layered space, closer to a Japanese woodblock print than a Renaissance corridor.</p><p>One caution: the broad categories are reliable, but the precise numbers are not. Models clearly tell wide from telephoto. They’re much less consistent at telling 70mm from 85mm.</p><h3>Aperture and the Two Meanings of Tilt-Shift</h3><p>Aperture is one of the most reliable camera terms you can use. Wide apertures, around f/1.2 to f/2, give shallow depth of field: a sharp subject against a soft, creamy background. Narrow apertures, around f/8 to f/16, keep everything sharp, the classic look for landscapes and architecture.</p><p>Effect words work just as well, and sometimes better: “shallow depth of field,” “bokeh,” “deep focus, everything sharp.” Thaddeus’s “background melting into soft blur” did the job of Sabine’s “f/1.4” without needing the number.</p><p>Avoid contradictions. “f/1.4” combined with “every detail of the background sharp” asks for two opposite things, and the model will compromise badly. In painting terms, aperture is about lost and found edges controlled through focus; decide which edges you want lost.</p><p>Tilt-shift is the term that trips people up, because it means two different things. To most models, it means the miniature effect: a thin band of focus that makes a real street look like a toy model. That’s the most visually distinctive use of the lens, so that’s what the model learned.</p><p>Real tilt-shift lenses are also used to straighten buildings so vertical lines don’t converge. Models rarely deliver that from the term alone. For straight architecture, describe the result instead: “architectural photography, corrected perspective, vertical lines parallel.”</p><p>There’s a general lesson here. When a term has two meanings, the model usually picks the one that is more visible.</p><h3>Angle and Viewpoint: The Most Underrated Controls</h3><p>If you learn only one group of camera terms, learn these. Viewpoint words reshape composition more reliably than any lens spec, because they describe something immediately visible in every photo.</p><p>“Overhead,” “top-down,” and “flat lay” put the camera directly above the subject. “Flat lay” is especially reliable for objects, food, and arrangements on a surface. “Bird’s-eye view” and “aerial view” open up the scale of landscapes and cities.</p><p>“Low angle” and “worm’s-eye view” make subjects feel powerful and monumental. “Eye level” keeps things neutral and intimate. A “Dutch angle,” with the camera tilted, adds unease.</p><p>Shot sizes borrowed from film also work well, because film stills are captioned with them so often: “extreme close-up,” “close-up,” “medium shot,” “full body,” “wide establishing shot.” These are often the fastest way to control how much of the subject fills the frame.</p><p>Thaddeus discovered this in his first week. Adding “seen from slightly below” to a portrait changed the composition more than every lens term Sabine tried.</p><h3>Shutter, Film Stocks, and Camera Bodies</h3><p>For shutter speed, effect words beat numbers almost every time. “Long exposure, light trails, silky water” gets you motion blur far more reliably than “2-second exposure.” “Frozen motion, droplets suspended in the air” beats “1/8000s.”</p><p>Film stocks are the surprise. Famous stocks have distinctive color and grain, and photographers name them in captions constantly, so the associations are strong. Portra tends toward warm, soft skin tones, Tri-X toward gritty black and white, and Velvia toward saturated landscapes. Treat these as tendencies to test, not guarantees.</p><p>Camera bodies are weaker. “Shot on Hasselblad” may nudge an image toward a polished, high-end look, and “Leica” toward a certain street-photography feel, but the effect is inconsistent. It’s mostly flavor rather than control.</p><p>Then there are the terms that rarely do anything visible: ISO values, white balance in Kelvin, exact shutter fractions, and long strings of specs. They cost words and dilute the terms that do work.</p><p>Sabine cut her spec sheet to three terms: one focal range, one aperture, and one film stock. Her portraits improved immediately.</p><h3>A Testing Method You Can Repeat</h3><p>Don’t take my word, or anyone’s, for which terms work in your tool. Test them. Every model learned slightly different associations, and they shift between versions.</p><p>Fix everything except the term you’re testing: the same prompt, the same aspect ratio, and the same seed if your tool lets you set one. Then change a single camera term and generate a small batch.</p><p>Lay the results out in a grid. A row of 24mm, 50mm, 85mm, and 200mm versions of one portrait will teach you more in five minutes than any article, including this one.</p><p>Keep the grids as a personal reference sheet for each tool and model version. Re-test the important terms whenever a model updates. A term that worked last spring may now do something different, or nothing at all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Snck1V7h31Ara5A80ZiBKA.png" /></figure><h3>A Starting Cheat Sheet</h3><p>Use this as a first guess, then confirm it with your own grids.</p><p><strong>Reliable:</strong> wide versus telephoto, shallow versus deep focus, viewpoint and camera angle, shot size, long exposure, and macro.</p><p><strong>Sometimes:</strong> specific focal lengths, film stocks, and the tilt-shift miniature effect.</p><p><strong>Rarely:</strong> ISO, exact shutter speeds, white balance values, and most camera body names.</p><p>Sabine still thinks like a photographer, but she now prompts like a painter: she names the effect first and adds a camera term only when it says the same thing in fewer words. Thaddeus picked up three camera terms that each saved him a sentence.</p><p>The model doesn’t know optics. It knows what photographs look like. Prompt for the look you want, and let camera language carry only the terms that have earned their place.</p><p><em>Jean-Marie Bonthous (publishing as J.M. Bonthous) is the retired founder of a successful Digital Marketing agency. He is the author or more than two dozen books, including six on message-driven marketing, seven on the human dimensions on AI and several on AI art and AI filmmaking</em>. <em>He writes four blogs on Medium — About the human dimensions of AI:</em> <a href="https://medium.com/@jm_26203"><em>AI in Real Life</em></a><em> About AI art:</em> <a href="https://medium.com/@jmbonthous"><em>The Algorithmic Eye</em></a><em> About AI filmmaking:</em> <a href="https://medium.com/@info_13818"><em>The Solitary Frame</em></a><em> About Message-Driven Marketing: </em><a href="https://medium.com/@jm_b_73712"><em>The Message-Driven Marketer</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abc93a1915c7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Renaissance Workshops to AI Labs: The Long History of Collaborative Art]]></title>
            <link>https://medium.com/@jmbonthous/from-renaissance-workshops-to-ai-labs-the-long-history-of-collaborative-art-4404f3be70d8?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4404f3be70d8</guid>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[digital-art]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[art]]></category>
            <category><![CDATA[artist]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 15:01:04 GMT</pubDate>
            <atom:updated>2026-10-05T15:01:04.070Z</atom:updated>
            <content:encoded><![CDATA[<p>Every artist has heard some version of the myth. The lone genius, alone in the studio, waiting for a flash of inspiration to strike. No assistants, no committee, no outside hands. Just one person and the work.</p><p>It’s a great story. It’s also almost never how art actually gets made.</p><p>Rembrandt ran a workshop with dozens of pupils grinding pigments, stretching canvases, and painting entire backgrounds before he ever picked up a brush. Some paintings that hang in museums under his name were mostly executed by his students, with Rembrandt adding a few final strokes and his signature.</p><p>Nobody calls those fakes. We call them Rembrandts.</p><p>If you make AI art and feel a flicker of guilt about “cheating,” or you’ve been told your Midjourney prints aren’t “real art” because a machine did some of the work, this history is worth knowing. You’re not the first artist to direct more than you personally execute. You’re just the latest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eMN7t4j2VE4dDaKgou8n5A.png" /></figure><h3><strong>The Renaissance Workshop Was a Machine With People In It</strong></h3><p>Michelangelo, Raphael, Titian — none of them worked alone. Renaissance workshops, called <em>bottegas</em>, operated like small factories. A master took commissions, set the vision, and assigned specific tasks to specialists: one apprentice for drapery, another for hands, another for backgrounds.</p><p>Raphael’s workshop in Rome employed around fifty assistants at its peak. Many of the frescoes attributed to him were painted largely by his team, following his drawings and under his direction. He was the creative director. They were the production crew.</p><p>Attribution back then was never about “who touched the canvas.” It was about whose vision organized the work. A patron commissioning a Raphael wanted Raphael’s eye, his composition, his sense of color — not proof that his hand alone moved the brush.</p><p>Even the raw materials weren’t solo achievements. Pigment makers ground lapis lazuli into ultramarine blue, a color so expensive it was often reserved for the Virgin Mary’s robes. Panel makers built the wood supports. Framers finished the object. The artist sat at the center of a supply chain, not outside of it.</p><p>This is worth sitting with if you’re an AI artist today. The “purity” people demand of solo, hand-made art was largely invented after the fact — mostly in the 19th century, when Romantic writers needed a mythology to match their idea of the tortured, singular genius.</p><h3><strong>Film Never Pretended to Be a One-Person Job</strong></h3><p>Jump forward a few centuries and look at film, an art form nobody has ever seriously claimed one person makes alone. A single movie credits a director, cinematographer, editor, production designer, composer, sound team, and often hundreds of visual effects artists.</p><p><em>The Lord of the Rings</em> trilogy credits over 2,000 people just in visual effects work across its productions. Peter Jackson is the “author” of those films in the same sense Raphael was the author of his frescoes — the vision was his, the execution was collective.</p><p>Auteur theory, the idea that a director’s personal vision unifies a film despite hundreds of contributing hands, is really just the genius myth adapted for an industrial art form. It let critics keep talking about “a Kubrick film” or “a Spielberg film” without pretending Kubrick operated the camera himself.</p><p>Audiences accepted this without much fuss. Nobody watches <em>Jaws</em> and feels cheated that Spielberg didn’t personally build the mechanical shark. We understand that directing a vision through other hands and tools is its own serious skill.</p><p>That distinction — between authorship as vision-and-direction versus authorship as literal manual execution — is exactly the distinction AI artists need today, and exactly the one most critics haven’t caught up to yet.</p><h3><strong>Game Studios Already Blurred Human and Machine Contribution</strong></h3><p>Before AI art, video games and animation had already normalized a strange kind of authorship: part human decision, part machine output. A single frame in a Pixar film passes through concept artists, riggers, lighting technicians, and render engines running physics simulations no individual person could calculate by hand.</p><p>Procedural generation in games like <em>No Man’s Sky</em> creates entire planets algorithmically, based on rules human designers wrote but didn’t personally draw. The designers didn’t paint each of the game’s 18 quintillion planets. They built a system that paints them, and then curated what the system produced.</p><p>Sound familiar? That’s already close to how a diffusion model works. You’re not describing every pixel. You’re building — or in your case, directing — a system, and then curating and refining what it gives back.</p><p>Animation pipelines went through the exact same anxiety AI art is going through now. When studios moved from hand-drawn cels to computer rendering in the 1990s, traditionalists said it wasn’t “real” animation anymore. Today nobody argues Pixar films aren’t real animation.</p><h3><strong>What an AI Art “Studio” Actually Looks Like</strong></h3><p>Here’s the practical version, stripped of the theory: when you make an image with Midjourney, you’re not working alone any more than Raphael was. You’re the director in a workshop with three other collaborators — the model, the dataset, and the algorithm.</p><p>The model brings trained visual instincts, shaped by exposure to millions of images, that will push toward certain compositions and lighting whether you ask for them or not. The dataset is the accumulated, uncredited labor of every photographer and painter whose work shaped those instincts. The algorithm — diffusion, sampling, upscaling — is the mechanical process turning your direction into pixels, the same role the panel-maker or the render engine played before.</p><p>In my own practice building a series of AI-generated documentary-style photographs, most of the real work isn’t typing a prompt and hoping. It’s building and refining a prompt system over dozens of iterations — a working vocabulary of lighting, film stock references, and compositional cues that reliably produces a specific humanist, black-and-white sensibility, closer to Doisneau or Álvarez Bravo than to generic AI output.</p><p>That system took months to develop, version by version, the same way a workshop master develops house techniques his apprentices learn to execute. The model contributes visual instinct. I contribute the eye, the editing, the rejection of the ninety images that don’t work for the one that does.</p><h3><strong>Curation Is Not a Lesser Kind of Authorship</strong></h3><p>If you’re an artist working with AI tools and someone tells you “you didn’t make that, the computer did,” here’s your answer: curation and direction have always been forms of authorship, not lesser substitutes for it.</p><p>Raphael didn’t personally paint every square inch of his frescoes, and nobody thinks less of the work for it. Spielberg didn’t build the shark. The Pixar animator didn’t calculate the physics of falling water by hand. Selection, direction, and refinement are the actual skill, and they always have been.</p><p>What’s different with AI isn’t the presence of collaborators — every serious art form has had those. What’s different is that one of your collaborators is a statistical system instead of a human apprentice, which makes people uncomfortable in a way that a paintbrush or a render engine never did.</p><p>That discomfort is worth naming honestly rather than arguing around. It’s fair to ask questions about the datasets AI models are trained on, about consent and compensation for the artists whose work shaped those models. Those are real, unresolved issues — not reasons to deny that direction and curation are genuine artistic labor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I4GSZuHn7ldrXd-1pw_LZQ.png" /></figure><h3><strong>The Workshop Never Closed</strong></h3><p>Picture an apprentice in Raphael’s studio, grinding lapis lazuli into blue pigment. Picture a Pixar technician tuning a physics simulation for how water should fall. Picture a diffusion model, trained on millions of images, generating a draft you’ll spend the next hour refining into something only you would have made.</p><p>Different centuries, different tools, same basic shape: a vision, directed through collaborators — some human, some material, some now computational — into a finished piece.</p><p>The myth of the solitary genius was never true. AI art doesn’t break five hundred years of tradition. It just makes the collaboration too obvious to keep pretending it wasn’t always there.</p><p><em>Jean-Marie Bonthous (publishing as J.M. Bonthous) is the retired founder of a successful Digital Marketing agency. He is the author or more than two dozen books, including six on message-driven marketing, seven on the human dimensions on AI and several on AI art and AI filmmaking</em>. <em>He writes four blogs on Medium — About the human dimensions of AI:</em> <a href="https://medium.com/@jm_26203"><em>AI in Real Life</em></a><em> About AI art:</em> <a href="https://medium.com/@jmbonthous"><em>The Algorithmic Eye</em></a><em> About AI filmmaking:</em> <a href="https://medium.com/@info_13818"><em>The Solitary Frame</em></a><em> About Message-Driven Marketing: </em><a href="https://medium.com/@jm_b_73712"><em>The Message-Driven Marketer</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4404f3be70d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Craft to Curation — What Does an Artist Do When the Machine Can Make Almost Anything?]]></title>
            <link>https://medium.com/@jmbonthous/from-craft-to-curation-what-does-an-artist-do-when-the-machine-can-make-almost-anything-3af051360e3e?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3af051360e3e</guid>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-art-gallery]]></category>
            <category><![CDATA[ai-artist]]></category>
            <category><![CDATA[ai-art-generator]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 15:01:03 GMT</pubDate>
            <atom:updated>2026-09-28T15:01:03.556Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine two artists sitting in front of exactly the same AI image generator.</p><p>Mara Voss has become extremely good at prompting. She understands how the model responds to different lenses, lighting instructions, stylistic references and compositional cues. By the end of an afternoon, she has generated 600 images.</p><p>Her best twelve are spectacular.</p><p>Across town, Ivo Serrin has spent the same afternoon with the same technology. He has generated nearly as many images, but most of his time has been spent looking rather than generating. He deletes hundreds. He returns repeatedly to a few that bother him in interesting ways.</p><p>At six o’clock, Mara has twelve beautiful images.</p><p>Ivo has three.</p><p>And strangely, Ivo may have done more artistic work.</p><p>Not because producing fewer images is inherently more creative. And certainly not because technical skill no longer matters.</p><p>Something else has happened.</p><p>Ivo’s primary creative act has shifted from making to <strong>judging</strong>.</p><p>That shift may turn out to be one of the most important consequences of generative AI for artists.</p><p><strong>When making becomes easy, choosing becomes hard.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XjSzEaMvyWgEBa3rSniMNw.jpeg" /></figure><h3>What Exactly Did the Artist Make?</h3><p>This question would have sounded peculiar during much of art history.</p><p>A painter painted. A sculptor sculpted. A draftsman drew. The connection between the artist’s hand and the resulting object seemed obvious.</p><p>Generative AI disrupts that connection.</p><p>I can describe an elderly woman sitting alone in a café in Marseille and, within minutes, receive dozens of plausible photographs of a woman who never existed, sitting in cafés that never existed, photographed by cameras that were never there.</p><p>So what exactly did I make?</p><p>I did not position the woman. I did not choose the actual light falling through the window. I did not press a shutter. In the traditional sense, I did not execute the image.</p><p>This has led critics of AI art to an understandable conclusion: the machine made it.</p><p>But that answer becomes less satisfactory once we watch serious AI artists work.</p><p>They conceive, direct, generate, reject, regenerate, combine, crop, modify, sequence and sometimes physically alter their images. Recent research on AI-mediated creativity describes precisely this redistribution of labor: away from direct execution and toward supervision, iteration, selection and interpretive framing.</p><p>The artist has not necessarily disappeared.</p><p>The artist may have changed jobs.</p><h3>Photography Already Changed the Job</h3><p>We have been through a version of this before.</p><p>When photography appeared in the nineteenth century, one objection was obvious. The photographer did not actually make the thing depicted.</p><p>A painter created every mark representing a tree. The photographer merely pointed a machine at a tree that already existed.</p><p>But this profoundly underestimated what photographers actually did.</p><p>Think about Henri Cartier-Bresson.</p><p>He did not create the man jumping across the puddle behind the Gare Saint-Lazare in 1932. He did not build the ladder lying in the water. He did not position the man’s reflection beneath him.</p><p>His artistic act was partly one of recognition.</p><p>Among the infinite number of visual arrangements occurring around him, he recognized the fraction of a second that mattered.</p><p>Photography introduced something fundamental into our understanding of artistic authorship: <strong>selection can itself be creative.</strong></p><h3>Look at the Contact Sheet</h3><p>Perhaps the best evidence is the photographer’s contact sheet.</p><p>In the film era, a photographer shooting 35mm film typically had 36 exposures on a roll. The contact sheet placed all those frames together so that the photographer could study them, mark possibilities and decide which ones deserved to become photographs.</p><p>The famous <em>Magnum Contact Sheets</em> project makes this process unusually visible. It brings together 139 contact sheets from 69 photographers across more than 70 years, including Cartier-Bresson, Elliott Erwitt and Inge Morath. What we see is not simply how photographs were taken, but how photographers edited their own seeing.</p><p>One frame survives. Thirty-five disappear.</p><p>Was rejecting the other 35 part of making the photograph?</p><p>Of course it was.</p><p>The contact sheet reveals something that the finished print conceals. Great photographers did not simply produce great photographs. They produced failures, near misses, repetitions and alternatives.</p><p>Then they recognized the one that mattered.</p><p>AI makes this old photographic process enormous.</p><p>Instead of 36 frames, I may have 360.</p><p>Or 3,600.</p><p>The contact sheet has become almost infinite.</p><h3>Ivo’s 347th Image</h3><p>Return to Ivo Serrin.</p><p>He is creating a series about people waiting for someone who will never arrive.</p><p>The machine generates beautiful railway stations, rain-soaked bus shelters, deserted airport terminals and solitary figures beneath streetlights.</p><p>Image 347 is technically unimpressive.</p><p>The man’s face is barely visible. Part of his body is obscured by another passenger. The station is less dramatic than in several earlier generations. There is an awkward empty space on the right.</p><p>Yet Ivo stops.</p><p>Something about the man’s posture suggests resignation rather than loneliness. The empty space suddenly feels necessary. The partially obscured face prevents the image from becoming sentimental.</p><p>Ivo saves it.</p><p>Why?</p><p>The AI cannot answer that question.</p><p>It generated image 347 just as willingly as it generated image 346 and image 348.</p><p>The significance was supplied by the person who recognized it.</p><h3>Is Curation Really Creation?</h3><p>This is where some people will object.</p><p>Surely choosing an image is not equivalent to painting one.</p><p>I agree.</p><p>They are different creative acts.</p><p>But we should be careful about assuming that physical execution is the only serious form of authorship.</p><p>Consider filmmaking.</p><p>Federico Fellini did not personally build every set in <em>8½</em>. He did not sew the costumes, operate the camera throughout the production, process the film or play Nino Rota’s score.</p><p>Hundreds of people contributed skills Fellini himself may not have possessed.</p><p>Yet calling <em>8½</em> a Fellini film makes perfect sense.</p><p>Why?</p><p>Because authorship can reside in orchestration.</p><p>The director establishes a vision, makes thousands of decisions, accepts some contributions and rejects others, asks for changes, shapes performances and ultimately determines what belongs inside the finished work.</p><p>Architecture operates similarly. The architect does not normally lay the bricks.</p><p>The conductor does not play the violin.</p><p>Creative authorship has never depended entirely upon personally executing every component of a work.</p><h3>Then Came Duchamp</h3><p>Art itself confronted the issue even more directly more than a century ago.</p><p>In 1917, Marcel Duchamp famously submitted an ordinary manufactured urinal to an exhibition, signed “R. Mutt,” and called it <em>Fountain</em>.</p><p>Duchamp had not fabricated the object.</p><p>He selected it.</p><p>Whatever one thinks of <em>Fountain</em> — and people have been arguing about it for more than a century — it forced art to confront an uncomfortable proposition.</p><p>Can choosing something, placing it in a new context and asking us to look at it differently constitute an artistic act?</p><p>AI gives that old question astonishing new relevance.</p><p>The difference is scale.</p><p>Duchamp selected one object from the physical world.</p><p>An AI artist can now select from a practically limitless world of possibilities that did not exist until they were generated.</p><h3>Mara Discovers That Her Skill Has Become a Button</h3><p>Meanwhile, Mara has a problem.</p><p>Six months earlier she learned an elaborate prompting technique that allowed her to achieve a particular cinematic lighting effect. It took weeks of experimentation.</p><p>Then the AI company releases a new model.</p><p>The same effect can now be achieved almost automatically.</p><p>Mara experiences something increasingly familiar to people working seriously with generative AI:</p><p><strong>A skill she worked hard to acquire has become a button.</strong></p><p>This is one of the peculiarities of AI art.</p><p>Technical AI skills can depreciate remarkably quickly.</p><p>Prompt syntax that seemed essential two years ago becomes irrelevant. Workarounds disappear. Models learn to understand ordinary language. Features requiring complicated workflows become menu choices.</p><p>That does not mean knowledge becomes worthless.</p><p>It means technical proficiency alone provides a fragile artistic advantage.</p><p>Judgment depreciates much more slowly.</p><h3>Taste Becomes a Serious Skill</h3><p>The word “taste” can sound superficial. It shouldn’t.</p><p>Taste is the accumulated ability to distinguish what is merely competent from what is interesting.</p><p>It tells the photographer that the technically imperfect frame is better.</p><p>It tells the editor that a beautiful paragraph should be deleted because it does not belong.</p><p>It tells the filmmaker that the quieter performance is more powerful.</p><p>And it tells the AI artist that image 347 is worth keeping while 346 is not.</p><p>This is one reason recent scholarship on generative creativity increasingly emphasizes human evaluation and selection. A 2026 study of AI-mediated cultural production concluded that generative systems expand the range of candidate outputs, while human creators remain responsible for interpreting, curating and turning those possibilities into culturally meaningful work.</p><p>Another recent account describes the emerging human role as that of a “prompter-curator”: framing the problem, selecting tools, critically evaluating results and taking responsibility for what ultimately gets presented.</p><p>That sounds much closer to what I see happening in serious AI art.</p><h3>But Curation Can Become an Excuse</h3><p>There is a danger in this argument.</p><p>If we define artistic authorship too generously, almost anything becomes art.</p><p>I type “beautiful old woman in Paris.”</p><p>The machine gives me four pictures.</p><p>I choose number three.</p><p>Am I now Cartier-Bresson?</p><p>Obviously not.</p><p>Clicking Generate fifty times and selecting the prettiest result does not automatically constitute meaningful artistic authorship.</p><p>Curation is only as interesting as the judgment behind it.</p><p>Why did I choose this image?</p><p>What did I reject?</p><p>What am I trying to say?</p><p>What does this image contribute to the other twenty images in the series?</p><p>Would another person with the same model, the same prompt and the same 500 outputs make the same choices?</p><p>The greater the role of curation, the more demanding those questions become.</p><h3>What You Reject May Define You</h3><p>This brings us to something I find particularly interesting about AI art.</p><p>We usually define artists by what they show us.</p><p>Perhaps we should also define them by what they refuse to show us.</p><p>Imagine that Mara and Ivo receive exactly the same 1,000 generated images.</p><p>Mara consistently rejects sentimentality.</p><p>Ivo rejects spectacle.</p><p>Another artist rejects beautiful faces because she finds ordinary ones more interesting. Another keeps compositions containing awkwardness and uncertainty. Another rejects every image that explains itself too easily.</p><p>After thousands of such decisions, patterns emerge.</p><p>Those patterns are not produced by the model.</p><p>They constitute taste.</p><p>Eventually they become something closer to a point of view.</p><p>And a point of view is one of the things we have traditionally expected artists to possess.</p><h3>The Artist as Editor</h3><p>There is another profession that may provide an even better analogy than curator.</p><p>The editor.</p><p>A film editor receives hours of footage and constructs minutes.</p><p>A book editor can transform a manuscript without writing most of its sentences.</p><p>A photography editor may discover relationships between pictures that were made months apart.</p><p>The creative act lies partly in exclusion.</p><p>This becomes especially important when AI removes scarcity from production.</p><p>When generating another image costs almost nothing, there is little pressure to stop.</p><p>We can always make another.</p><p>And another.</p><p>And another.</p><p>The artist’s problem therefore changes.</p><p>Scarcity once existed at the point of production.</p><p>AI moves scarcity toward attention and judgment.</p><p>The difficult act is no longer necessarily producing another possibility.</p><p>It is deciding when possibilities have become noise.</p><h3>Even Copyright Law Is Noticing the Difference</h3><p>Interestingly, the distinction between generation and creative selection has begun appearing in copyright policy.</p><p>The U.S. Copyright Office concluded in its 2025 report on AI-generated works that merely providing prompts does not, by itself, necessarily provide sufficient human authorship for copyright protection.</p><p>But the Office explicitly recognized something else.</p><p>A person may contribute protectable authorship by creatively <strong>selecting, coordinating, arranging or modifying</strong> AI-generated material.</p><p>Copyright law and artistic merit are certainly not the same thing.</p><p>Still, the distinction is revealing.</p><p>Even the law is beginning to confront the possibility that the important human contribution may sometimes occur after generation.</p><h3>Mara Changes Her Mind</h3><p>Several months later, Mara’s work looks different.</p><p>She still knows how to prompt beautifully. She has not abandoned technical skill.</p><p>But she has stopped trying to make impressive individual images.</p><p>Instead, she is working on a series about the apartment where her grandmother lived during the final years of her life.</p><p>She generates hundreds of rooms.</p><p>Most are gorgeous.</p><p>She rejects them.</p><p>The wallpaper is too picturesque. The sunlight is too cinematic. The old woman’s face is too expressive. One image looks like an advertisement for an expensive European perfume.</p><p>Then she finds something quieter.</p><p>An unmade bed.</p><p>A glass of water.</p><p>A slightly open door.</p><p>Nothing spectacular happens.</p><p>She keeps it.</p><p>Mara has begun doing what Ivo was doing from the beginning.</p><p>She is no longer asking the machine what beautiful things it can produce.</p><p>She is deciding what belongs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9NPI7pyrH_xHyUhbC4TLnA.jpeg" /></figure><h3>From Craft to Curation</h3><p>None of this means craft disappears.</p><p>Painters will still paint. Photographers will still photograph. Some AI artists will develop highly technical workflows requiring considerable skill.</p><p>But our definition of artistic skill is expanding.</p><p>The artist increasingly becomes an originator, director, selector, editor and sequencer as well as a maker.</p><p>This connects with several changes already occurring across AI art.</p><p>When technical perfection becomes cheap, the human hand becomes valuable.</p><p>When generic imagery becomes abundant, personal experience becomes valuable.</p><p>When complete control becomes easy, accident and imperfection become valuable.</p><p>And when production becomes easy, <strong>judgment becomes valuable.</strong></p><p>Perhaps that is the larger transition we are witnessing.</p><p>For centuries, we often recognized artists by what their hands could do.</p><p>AI complicates that definition because the hand may no longer execute every part of the work.</p><p>But perhaps we have confused execution with authorship.</p><p>A photographer does not create the street. A director does not personally construct every frame. An architect does not lay the bricks. Their authorship resides partly in deciding what should exist, recognizing what works, rejecting what does not and imposing coherence on thousands of possibilities.</p><p>AI gives the artist something unprecedented: an almost unlimited supply of possibilities.</p><p>That sounds like liberation.</p><p>It can also become a burden.</p><p>Because when almost anything can be made, the artist’s most important question is no longer:</p><p><strong>Can I make this?</strong></p><p>It becomes:</p><p><strong>Of everything I could make, what is worth making?</strong></p><p>And answering that question is not a technical skill.</p><p>It is judgment.</p><p><strong><em>Jean Marie Bonthous</em></strong><em> (publishing as JM Bonthous) is the author of more than two dozen books, including six on the human side of AI, six about filmmaking, and four about digital/AI art. See his latest books: </em><a href="http://www.jmbonthous.com/">www.jmbonthous.com</a> <em>He writes three blogs on Medium: About the human dimensions of AI: </em><a href="https://medium.com/@jm_26203">AI in Real Life</a> <em>About AI art: </em><a href="https://medium.com/@jmbonthous">The Algorithmic Eye</a><em>About AI filmmaking: </em><a href="https://medium.com/@info_13818">The Solitary Frame</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3af051360e3e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why “Fair Use” Isn’t a Clear Answer for Visual Artists]]></title>
            <link>https://medium.com/@jmbonthous/why-fair-use-isnt-a-clear-answer-for-visual-artists-c5c848d41ccf?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5c848d41ccf</guid>
            <category><![CDATA[ai-art-gallery]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[ai-artist]]></category>
            <category><![CDATA[ai-art-generator]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 15:06:01 GMT</pubDate>
            <atom:updated>2026-09-21T15:06:01.366Z</atom:updated>
            <content:encoded><![CDATA[<p>Spend enough time around AI art and sooner or later someone will offer a reassuring answer to the copyright problem.</p><p>“It’s fair use.”</p><p>The phrase has become almost magical. AI companies use copyrighted material to train models? Fair use. An artist uses another artist’s work as a reference? Fair use. An AI-generated image resembles an existing photograph? Probably transformative, therefore fair use.</p><p>If only copyright law were that simple.</p><p>Fair use is real, and it provides important protection for artists, journalists, teachers, researchers, critics and others who legitimately build upon existing culture. But it is not a blanket exemption from copyright.</p><p>More importantly for AI artists, it is not a permission slip handed out in advance.</p><p><strong>Fair use is a legal argument you may have to make after somebody claims you infringed their copyright.</strong></p><p>And whether that argument succeeds depends heavily on what you actually did.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FBeoOD5Xfsh8aSBb__AX4w.png" /></figure><h3>Four Questions, Not One Rule</h3><p>American copyright law identifies four factors courts consider when deciding whether an unauthorized use of copyrighted material qualifies as fair use.</p><p>The first considers the purpose and character of the new use, including its commercial nature and whether it has a sufficiently different purpose or character. The second looks at the nature of the copyrighted work. The third considers how much of the original was taken and how significant that portion was.</p><p>The fourth asks about the effect on the market for the original work.</p><p>These are not four boxes where checking three means you win.</p><p>Courts weigh them together. Different circumstances can produce different answers. Even the same action can look different depending upon why it was done and what economic effect it produces.</p><p>That ambiguity existed long before generative AI.</p><p>AI has made it considerably more complicated.</p><h3>Consider a Photograph</h3><p>Imagine that I take one of my street photographs with my Ricoh GR III.</p><p>It shows an elderly man sitting alone at a café table. The composition, expression, background and light happen to come together in a way that makes the photograph distinctive.</p><p>Another artist downloads it.</p><p>What happens next matters enormously.</p><p>Suppose she writes an article criticizing contemporary street photography and reproduces my photograph to explain what she dislikes about my composition. That is the sort of critical use for which fair use has traditionally provided considerable breathing room.</p><p>Now change the facts.</p><p>She prints my photograph on T-shirts and sells them.</p><p>The fair-use argument becomes considerably harder.</p><p>Change them again.</p><p>She uses my photograph as input to an AI system and creates something dramatically different from it.</p><p>Now we are in much less familiar territory.</p><p>How different is different enough?</p><p>There is no percentage meter in copyright law that answers that question.</p><h3>Transformation Helps, but It Is Not Magic</h3><p>For years, discussions of fair use frequently centered on whether a new work was “transformative.”</p><p>That word has consequently become another piece of magical vocabulary in AI circles.</p><p>“I transformed it” can sound like the end of the discussion.</p><p>It isn’t.</p><p>The Supreme Court reinforced this point in its 2023 decision involving Andy Warhol’s portrait of Prince. Warhol had unquestionably transformed Lynn Goldsmith’s original photograph aesthetically. His silkscreen looked very different from her photograph.</p><p>Yet the Court ruled against the Andy Warhol Foundation in the particular commercial licensing use before it.</p><p>Why?</p><p>One important consideration was that Goldsmith’s photograph and the Warhol image had been licensed for similar purposes: magazine illustrations depicting Prince.</p><p>Visual transformation alone did not settle the matter.</p><p>That should be sobering for AI artists.</p><p>Changing what something looks like is not necessarily the same as changing what it is being used for.</p><h3>Then AI Training Entered the Courtroom</h3><p>Now consider what happened in 2025.</p><p>In <em>Bartz v. Anthropic</em>, three authors sued Anthropic over the use of copyrighted books in connection with training its Claude language models.</p><p>Federal Judge William Alsup concluded that Anthropic’s use of books to train its models was highly transformative and qualified as fair use under the circumstances before him.</p><p>That sounds like the simple answer AI companies had been waiting for.</p><p>Except it wasn’t.</p><p>The same judge treated another part of Anthropic’s conduct very differently.</p><p>Anthropic had downloaded millions of books from pirate libraries while building a permanent digital collection. Judge Alsup concluded that those pirated library copies were not protected by fair use merely because some could later be useful for AI training.</p><p>Same company.</p><p>Same broad technological project.</p><p>Different copying.</p><p>Different fair-use answer.</p><p>That alone should tell us why the phrase “AI training is fair use” is too simple.</p><h3>And Another Court Said No</h3><p>A few months earlier, another federal court had reached an important result pointing in the opposite direction.</p><p>Thomson Reuters, owner of the Westlaw legal research service, sued Ross Intelligence after Ross used material derived from Westlaw’s copyrighted headnotes to help train a competing AI legal-research product.</p><p>The Delaware federal court rejected Ross’s fair-use defense.</p><p>The market question mattered considerably. Ross was developing a product intended to compete with Westlaw, and the court also considered a potential market for licensing Westlaw material as AI training data.</p><p>There is an important qualification here. Ross was not operating a generative AI system comparable to Midjourney or ChatGPT.</p><p>Still, the case makes an important point.</p><p>There is no legal principle saying:</p><p><strong>AI + copyrighted material = fair use.</strong></p><p>The facts matter.</p><h3>Training Is Also Different From Output</h3><p>This distinction is frequently lost in discussions among AI artists.</p><p>There are really several copyright questions hiding inside what people casually call “the AI copyright issue.”</p><p>Was copyrighted material legally obtained?</p><p>Was making a copy for training permissible?</p><p>What does the model retain from the training?</p><p>Can it reproduce protected elements of a particular work?</p><p>Does an output substantially resemble an existing copyrighted work?</p><p>Can the person who generated the new image claim copyright in it?</p><p>These are different questions.</p><p>A court could eventually conclude that an AI company’s training practices were lawful while still concluding that a particular image generated by that system infringed somebody’s copyright.</p><p>The legality of the oven does not automatically determine the legality of everything that comes out of it.</p><h3>What About “In the Style of”?</h3><p>Here we encounter one of the most confusing issues for visual artists.</p><p>Suppose I ask an AI system:</p><p>“Create a portrait in the style of Vincent van Gogh.”</p><p>Van Gogh died in 1890, so copyright in his paintings has long since expired.</p><p>Now suppose I replace Van Gogh with a living illustrator.</p><p>The ethical issue immediately becomes uncomfortable, but the copyright issue is surprisingly complicated.</p><p>Copyright generally protects particular creative expression, not an artistic style in the abstract. A living artist does not simply own every use of certain colors, brushstrokes, compositions or visual characteristics associated with his or her work.</p><p>That does not mean anything goes.</p><p>An output might reproduce protected elements of a particular copyrighted work closely enough to create an infringement problem. Other legal theories can also arise depending on how an artist’s name, identity or work is used.</p><p>But “that looks like my style” and “that infringes my copyright” are not automatically the same claim.</p><p>This is one reason the ethical question and the legal question should not be confused.</p><p>Something can be legally defensible and still be artistically parasitic.</p><h3>The Market Question May Become Enormous</h3><p>Of the four fair-use factors, the effect on the market may become particularly contentious in the age of generative AI.</p><p>Imagine an illustrator who spent twenty years developing a distinctive visual language.</p><p>A company trains an AI system on thousands of images, including her work. Users can then request illustrations resembling the kind of work clients previously hired her to produce.</p><p>What exactly is the relevant market harm?</p><p>Did the AI copy particular works?</p><p>Did it learn general stylistic characteristics?</p><p>Has it created a substitute for the artist’s existing images?</p><p>Or has it created a substitute for the artist herself?</p><p>Courts are only beginning to confront versions of these questions.</p><p>A 2026 analysis in the <em>Journal of the Copyright Society</em> identifies precisely this tension. Generative AI complicates the traditional relationship between transformation and market substitution because a system might not reproduce an original work while potentially extracting value from many copyrighted works to create competing material.</p><p>That is a much harder question than whether somebody photocopied a photograph.</p><h3>There Is Another Copyright Problem for AI Artists</h3><p>So far we have been talking about the copyright belonging to other people.</p><p>But AI artists face the opposite problem as well.</p><p>Who owns <em>our</em> images?</p><p>The U.S. Copyright Office addressed that question directly in 2025.</p><p>Its position is relatively straightforward: copyright requires human authorship. Using AI as part of the creative process does not prevent copyright protection, but purely AI-generated material is not automatically protected.</p><p>Human contributions can be.</p><p>If an artist substantially modifies AI-generated material, combines it creatively with human-authored elements, or exercises sufficient creative control over protected expressive elements, those human contributions may qualify for copyright. Merely writing prompts, according to the Copyright Office’s current position, is generally not enough.</p><p>This creates a fascinating asymmetry.</p><p>An AI artist can worry about infringing someone else’s copyright while simultaneously discovering that parts of his own AI-generated image may not receive copyright protection.</p><h3>What Should a Working AI Artist Actually Do?</h3><p>None of this means visual artists should become terrified of copyright law.</p><p>It means we should stop looking for one-sentence answers.</p><p>If I am creating an image inspired by hundreds of photographs I have seen during my life, that presents one set of questions.</p><p>If I upload a particular copyrighted photograph and ask AI to make a slightly altered version, that presents another.</p><p>If I generate something and later discover that it closely resembles an existing copyrighted image, I have another problem again.</p><p>And if I intend to sell the result, license it for advertising or put it on the cover of a book, the consequences become more significant.</p><p>The sensible question is not simply, “Is AI fair use?”</p><p>It is:</p><p><strong>What exactly am I using, what am I doing with it, how much am I taking, what am I producing, and could my new work substitute for the original?</strong></p><p>Those questions will not always provide a definitive answer.</p><p>They will provide a much better one.</p><h3>Law, Ethics and Artistic Identity</h3><p>There is also a larger issue here that interests me more than copyright doctrine itself.</p><p>An artist should not need a judge to establish the boundaries of his creative identity.</p><p>Suppose the law eventually gives AI artists enormous freedom to imitate styles, train systems on existing art and generate images derived from vast quantities of human creativity.</p><p>Would we necessarily want to use all that freedom?</p><p>Probably not.</p><p>The most interesting AI artists are increasingly trying to develop recognizable personal voices. They are introducing their own memories, experiences, imperfections and obsessions into a technology whose great temptation is to make everything available to everyone.</p><p>Copying another artist’s visual identity may therefore become less interesting precisely as it becomes easier.</p><p>The legal question asks:</p><p><strong>Can I do this?</strong></p><p>The artistic question asks:</p><p><strong>Why would I want to?</strong></p><p>The second question may ultimately matter more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mt127JGxYCsEuvBiTYRdXg.png" /></figure><h3>Fair Use Is a Conversation, Not a Permission Slip</h3><p>We are still remarkably early in this process.</p><p>The U.S. Copyright Office began its major AI initiative only in 2023 and received more than 10,000 public comments. Its reports are being released in stages as courts simultaneously work through dozens of AI-related copyright disputes.</p><p>Some courts have already found certain AI-training uses to be fair use.</p><p>Another has rejected fair use in an AI-related training dispute.</p><p>Future cases will involve different models, different source materials, different outputs and different evidence of market harm.</p><p>Eventually, clearer boundaries will emerge.</p><p>But visual artists working with AI cannot wait for ten years of appellate decisions before making pictures.</p><p>We have to work inside the ambiguity.</p><p>That means understanding what fair use actually offers us.</p><p>It is not a guarantee.</p><p>It is not a percentage.</p><p>It is not a declaration that anything “transformative” is automatically legal.</p><p>And it certainly is not a universal exemption for anything involving artificial intelligence.</p><p>Fair use is one of copyright law’s most important accommodations between protecting creators and allowing culture to build upon what came before.</p><p>Its ambiguity is partly what allows it to perform that function.</p><p>For visual artists working with AI, though, that ambiguity leaves us with an uncomfortable conclusion.</p><p><strong>“It’s fair use” is not the answer.</strong></p><p>Very often, it is where the question begins.</p><p><strong><em>Jean Marie Bonthous</em></strong><em> (publishing as JM Bonthous) is the author of more than two dozen books, including six on the human side of AI, six about filmmaking, and four about digital/AI art. See his latest books: </em><a href="http://www.jmbonthous.com/">www.jmbonthous.com</a> <em>He writes three blogs on Medium: About the human dimensions of AI: </em><a href="https://medium.com/@jm_26203">AI in Real Life</a> <em>About AI art: </em><a href="https://medium.com/@jmbonthous">The Algorithmic Eye</a> <em>About AI filmmaking: </em><a href="https://medium.com/@info_13818">The Solitary Frame</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5c848d41ccf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Authenticity Rebellion — Imperfection as Aesthetic]]></title>
            <link>https://medium.com/@jmbonthous/the-authenticity-rebellion-imperfection-as-aesthetic-a858d9fb9f2f?source=rss-9897f835057e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a858d9fb9f2f</guid>
            <category><![CDATA[ai-artist]]></category>
            <category><![CDATA[ai-art-gallery]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[ai-art-generator]]></category>
            <category><![CDATA[ai-art-prompts]]></category>
            <dc:creator><![CDATA[JM Bonthous]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 15:01:03 GMT</pubDate>
            <atom:updated>2026-09-14T15:01:03.684Z</atom:updated>
            <content:encoded><![CDATA[<p>For most of the history of photography, imperfection was something we tried to eliminate.</p><p>We wanted sharper lenses, finer-grained film, more accurate exposure and better control over focus. Digital photography accelerated the process. Noise reduction removed grain. Image stabilization reduced camera shake. Photoshop removed blemishes. Computational photography rescued shadows and highlights that photographers once simply lost.</p><p>Then generative AI arrived and offered something approaching the logical conclusion of this long pursuit: images in which practically everything could be controlled.</p><p>And almost immediately, artists started putting the mistakes back.</p><p>Film grain. Light leaks. Motion blur. Crooked horizons. Blown highlights. Awkward framing. Weathered skin. Stray hair. Unbalanced compositions. Objects intruding into the frame.</p><p>Something rather strange is happening in AI art.</p><p><strong>We spent decades teaching technology to eliminate photographic imperfections. Now we are teaching machines to recreate them.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IJ9D2vFSdnLYDKIRWOmjzQ.png" /></figure><h3>The Problem With the AI Look</h3><p>You probably recognize it even if you cannot explain precisely what is wrong.</p><p>The skin is slightly too smooth. The lighting is beautiful but seems to come from nowhere. The subject is positioned perfectly. The background complements the foreground. Every object appears to have been arranged by an invisible art director.</p><p>Nothing is necessarily wrong with the image.</p><p>And that is precisely what is wrong with it.</p><p>This is no longer merely an impression among people who spend too much time looking at AI imagery. A 2026 paper in <em>The Journal of Aesthetics and Art Criticism</em> examined what its authors describe as the aesthetic biases of generative AI systems. Even without explicit instructions, models frequently produce images resembling professionally staged photographs: centered subjects, balanced framing, neatly organized objects and carefully contrasted backgrounds.</p><p>The researchers noticed something else. Ordinary photographic imperfections such as minor motion blur, lens flare and camera shake tend to disappear unless specifically requested. Human faces can become unusually symmetrical. Lighting and shadows can look strangely idealized.</p><p>The authors describe the result as a replacement of the messiness of ordinary visual experience by a kind of visual perfection and order.</p><p>That is a remarkably good description of the AI look.</p><h3>We Are Learning to See AI</h3><p>There is another development taking place at the same time.</p><p>People are becoming accustomed to seeing synthetic images.</p><p>LTX Studio, in its analysis of 2026 AI image trends, describes audiences as increasingly sensitive to familiar AI characteristics: synthetic-looking skin, implausible lighting and compositions that seem suspiciously symmetrical. Its predicted response is a growing preference for natural skin texture, grain, light leaks, ordinary environments and images resembling moments that were discovered rather than manufactured.</p><p>The claim that everyone can now instantly detect AI would be an exaggeration. Research suggests that judging authenticity remains difficult. In one 2025 experiment involving 292 participants, more realistic AI-generated images were actually more likely to be judged authentic, although participants were less confident in those judgments.</p><p>But something subtler seems undeniable.</p><p>We are developing a visual vocabulary for artificial perfection.</p><p>Once we recognize it, perfection itself can become suspicious.</p><h3>Why My Ricoh GR III Still Matters</h3><p>This is one reason I remain fascinated by a very different kind of image-making: walking around with a small Ricoh GR III.</p><p>The Ricoh does not necessarily produce imperfect photographs. It is an extremely capable camera. But the way I use it encourages imperfection.</p><p>It can be pulled from a pocket in seconds. There is no viewfinder. There is often little time to compose. A person may already be leaving the frame. Someone else may walk into it. The light may be terrible. The horizon may not be straight.</p><p>Sometimes the photograph succeeds partly because of these things rather than in spite of them.</p><p>Street photography has always understood this.</p><p>Consider Robert Frank’s <em>The Americans</em>. Many of its photographs would fail the standards by which we judge technical perfection today. They can be grainy, dark, tilted and roughly framed. Faces disappear into shadows. Objects interrupt compositions.</p><p>Yet removing those characteristics would not improve the photographs.</p><p>It would destroy them.</p><p>The imperfections tell us something essential: <strong>this happened.</strong></p><h3>The Difference Between a Picture and a Moment</h3><p>That may be the real challenge facing AI photography.</p><p>AI is exceptionally good at making pictures.</p><p>Photography, at its best, often gives us moments.</p><p>The difference is difficult to define but easy to feel.</p><p>Look at photographs by Garry Winogrand. People are cut off by the frame. Bodies overlap inconveniently. Horizons tilt. Expressions are unresolved. The visual world refuses to arrange itself politely for the photographer.</p><p>Henri Cartier-Bresson could impose extraordinary formal order on chaos, but the chaos remained. The famous leap behind the Gare Saint-Lazare works because everything aligns for a fraction of a second. A moment later, the geometry disappears.</p><p>The photographer discovers order inside disorder.</p><p>AI often does the reverse.</p><p>It manufactures order and then asks us to believe disorder happened around it.</p><p>That difference matters.</p><h3>So We Tell AI to Make Mistakes</h3><p>The response from AI creators is fascinating.</p><p>Instead of asking for higher image quality, they increasingly ask for lower apparent image quality.</p><p>LTX’s own examples of prompts for realistic-looking people include 35mm film grain, heavy grain, halation, low contrast and unusual camera angles. Other current analyses of AI-image trends point to direct flash, blown highlights, hazy textures, light leaks, off-center framing, underexposure and analog color palettes as increasingly popular ways of escaping the polished AI aesthetic.</p><p>This represents an extraordinary reversal.</p><p>We used to use technology to correct the camera’s failures.</p><p>Now we use technology to simulate them.</p><p>But the most interesting AI artists can go further. Instead of merely adding a film-grain filter to an otherwise immaculate generation, they can encourage the system to misbehave.</p><p>Pixelation can become texture. Distortion can become expression. An incorrectly rendered space can become surrealism. Warped geometry can make an ordinary room unsettling. A figure partially dissolving into its surroundings may be far more interesting than a perfectly rendered person.</p><p>The mistake stops being a failure.</p><p>It becomes material.</p><h3>When the Accident Becomes the Art</h3><p>This idea is certainly not new.</p><p>Artists have been exploiting accidents for centuries.</p><p>Watercolor spreads. Ink bleeds. Paint drips. Photographic film gets fogged. Polaroid chemistry behaves unpredictably. Darkroom printing introduces dust, scratches and uneven exposure.</p><p>Jackson Pollock made controlled accident central to painting. Francis Bacon sometimes embraced distortions and unintended marks because they suggested possibilities more interesting than what he had consciously planned.</p><p>Photography has an especially rich relationship with failure.</p><p>William Klein embraced grain, blur and aggressive contrast. Daido Moriyama pushed those qualities even further. His photographs can be rough, dark, unstable and apparently careless. What would have been technical defects in conventional photography became a visual language.</p><p>AI offers a new variation on this old artistic strategy.</p><p>The machine makes different mistakes.</p><p>The artist’s job may be to recognize which ones are worth keeping.</p><h3>But Fake Grain Is Not Authenticity</h3><p>There is an obvious danger here.</p><p>If everyone begins prompting for grain, light leaks, crooked framing and imperfect skin, we will simply create another AI cliché.</p><p>The perfectly imperfect AI photograph may become as recognizable as the glossy AI photograph it replaced.</p><p>Indeed, this is already happening. A July 2026 account of the growing AI film-grain trend describes a cluster of increasingly familiar characteristics: visible grain, faded colors, light leaks, slight motion blur and imperfect framing.</p><p>Grain alone does not make an image authentic.</p><p>Nor does blur.</p><p>A fake scratch does not somehow imbue a synthetic photograph with a history it never had.</p><p>If anything, blindly applying these devices risks producing a new kind of artificiality: images carefully engineered to look accidental.</p><p>The goal cannot simply be to make AI look like film.</p><p>Something deeper has to happen.</p><h3>Imperfection Is Interesting When It Means Something</h3><p>The imperfections in a Robert Frank photograph matter because they are consequences of the circumstances in which the photograph was made.</p><p>The blur tells us somebody moved.</p><p>The grain tells us there was not much light.</p><p>The tilted frame tells us the photographer reacted quickly.</p><p>The partially obscured person reminds us that the world did not arrange itself for the camera.</p><p>Those imperfections contain information.</p><p>An AI artist therefore faces an interesting challenge. If our imperfections are generated rather than physically caused, they need to serve some other purpose.</p><p>Perhaps distortion expresses psychological unease. Perhaps excessive grain pushes an image toward memory rather than documentation. Perhaps an awkward crop creates tension. Perhaps an impossible shadow makes an otherwise ordinary domestic scene disturbing.</p><p>The imperfection needs a reason to exist.</p><p>Otherwise it is decoration.</p><h3>From Eliminating Errors to Curating Them</h3><p>This changes the role of the AI artist in a way I find particularly interesting.</p><p>The beginner looks at a generation and asks: <strong>What is wrong with it?</strong></p><p>The more mature artist may increasingly ask: <strong>Is what is wrong with it interesting?</strong></p><p>That is a very different creative question.</p><p>Suppose the model generates a woman’s face reflected incorrectly in a café window. The instinctive response is to regenerate it.</p><p>But perhaps the incorrect reflection makes the image mysterious.</p><p>Perhaps a hand has six fingers but appears only briefly at the edge of the frame. Does correcting it actually make the composition stronger?</p><p>Perhaps the architecture behind a subject could not exist in three-dimensional space. But perhaps that instability creates precisely the feeling the artist wants.</p><p>AI companies quite understandably work hard to eliminate these errors.</p><p>Artists do not necessarily have to cooperate.</p><h3>The Authenticity Paradox</h3><p>We arrive at a wonderful contradiction.</p><p>AI images may begin to feel more authentic precisely because artists stop asking AI to behave perfectly.</p><p>That does not make them authentic photographs. An invented moment remains invented. A simulated light leak does not mean light accidentally entered a camera.</p><p>But artistic authenticity is not the same thing as documentary authenticity.</p><p>An image can be completely fictional and still express something genuine.</p><p>The rebellion against AI perfection is therefore not really about fooling viewers into thinking an AI image was shot on film.</p><p>It is about rejecting the machine’s default idea of beauty.</p><p>Perfect skin is not necessarily beautiful. Perfect symmetry is not necessarily interesting. Perfect exposure does not necessarily produce atmosphere. Perfect composition can become lifeless.</p><p>Sometimes a picture needs something to go wrong.</p><h3>Teaching the Machine to Lose Control</h3><p>The first generation of AI image makers spent enormous amounts of time learning how to control the machine.</p><p>That made sense. Early systems were unpredictable, and technical mastery meant learning how to make them obey.</p><p>Now the more interesting challenge may be learning when not to make them obey.</p><p>Leave something unresolved.</p><p>Accept the strange reflection.</p><p>Let the face disappear into shadow.</p><p>Keep the motion blur.</p><p>Allow something awkward to intrude into the frame.</p><p>Generate the visual equivalent of a camera pulled from a pocket half a second too late.</p><p>This connects the Authenticity Rebellion to the larger changes taking place across AI art. In Post-AI Expressionism, artists restore the human hand. In personal storytelling, they restore the human life.</p><p>Here they restore something else that generative technology has been extraordinarily good at eliminating.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JCGvpUPh_b70qZsb-5qCrQ.png" /></figure><h3><strong>They restore chance.</strong></h3><p>For years, the great promise of AI image generation was control. We could finally create precisely the image we imagined.</p><p>We are now discovering the limitation of that promise.</p><p>Real life does not have perfect lighting. People do not stand in perfect compositions. Faces are not symmetrical. Streets are cluttered. People blink. Cameras move. Light leaks through cracks. Strangers walk into the frame.</p><p>The world does not collaborate with the photographer.</p><p>And perhaps that resistance is part of what made photography interesting in the first place.</p><p>The next generation of AI artists may therefore find themselves doing something wonderfully paradoxical.</p><p>After acquiring a machine capable of creating almost anything perfectly, they may spend much of their time teaching it how to fail.</p><p><strong><em>Jean Marie Bonthous</em></strong><em> (publishing as JM Bonthous) is the author of more than two dozen books, including six on the human side of AI, six about filmmaking, and four about digital/AI art. See his latest books: </em><a href="http://www.jmbonthous.com/">www.jmbonthous.com</a> <em>He writes three blogs on Medium: About the human dimensions of AI: </em><a href="https://medium.com/@jm_26203">AI in Real Life</a> <em>About AI art: </em><a href="https://medium.com/@jmbonthous">The Algorithmic Eye</a> <em>About AI filmmaking: </em><a href="https://medium.com/@info_13818">The Solitary Frame</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a858d9fb9f2f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>