<?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 Mario Lee on Medium]]></title>
        <description><![CDATA[Stories by Mario Lee on Medium]]></description>
        <link>https://medium.com/@qsibmini123?source=rss-7f85e88692f0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*d8mO8zcKiJMyjAuw</url>
            <title>Stories by Mario Lee on Medium</title>
            <link>https://medium.com/@qsibmini123?source=rss-7f85e88692f0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:45:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@qsibmini123/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[Why 3D Gaussian Splatting Is a Memory Problem — And Why HBM Matters for NVIDIA]]></title>
            <link>https://medium.com/@qsibmini123/why-3d-gaussian-splatting-is-a-memory-problem-and-why-hbm-matters-for-nvidia-f7c62e5a0355?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/f7c62e5a0355</guid>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[gpu-architecture]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[hbm]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 21 Dec 2025 14:03:48 GMT</pubDate>
            <atom:updated>2025-12-21T14:03:48.036Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Why 3D Gaussian Splatting Is a Memory Problem — And Why HBM Matters for NVIDIA</strong></h3><p>When people talk about <strong>3D Gaussian Splatting</strong>, they usually frame it as a rendering breakthrough.<br>Faster than NeRFs. Cleaner than point clouds. Easier than meshes.</p><p>But that framing misses the real bottleneck.</p><p><strong>Gaussian Splatting is not limited by graphics.<br>It’s limited by memory.</strong></p><p>And that single fact explains why NVIDIA dominates this space, why mobile devices struggle, and why future Physical AI systems will be memory-centric — not compute-centric.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Tpy4BhHTkWGjRsDQ" /></figure><h3>Gaussian Splats Don’t Stress the GPU — They Starve It</h3><p>A 3D Gaussian Splat scene isn’t one object.</p><p>It’s <strong>millions of tiny primitives</strong>, each carrying:</p><ul><li>position</li><li>covariance (scale + orientation)</li><li>opacity</li><li>color (often spherical harmonics)</li><li>sometimes motion or semantic attributes</li></ul><p>None of this math is especially hard.</p><p>The problem is <strong>feeding it to the GPU fast enough</strong>.</p><p>Each frame requires:</p><ul><li>fetching vast numbers of splats,</li><li>often with <strong>irregular, non-coalesced memory access</strong>,</li><li>with limited spatial or temporal reuse.</li></ul><p>In practice, the GPU isn’t busy computing.<br>It’s waiting.</p><p>That’s the textbook definition of a <strong>memory-bound workload</strong>.</p><h3>FLOPs Are Cheap. Bandwidth Is Not.</h3><p>Modern GPUs are overflowing with compute:</p><ul><li>CUDA cores</li><li>Tensor cores</li><li>specialized acceleration units</li></ul><p>But Gaussian Splatting barely benefits from more math.</p><p>What it really needs is:</p><ul><li>high sustained <strong>memory bandwidth</strong></li><li>low-latency access to large working sets</li><li>predictable delivery to keep pipelines full</li></ul><p>This is why:</p><ul><li>raw TFLOPs don’t correlate well with splat performance</li><li>shader optimization yields diminishing returns</li><li>cache misses dominate frame time</li></ul><h3>Why Meshes and Triangles Were Easier</h3><p>Traditional mesh-based graphics were fundamentally <strong>reuse-friendly</strong>:</p><ul><li>Vertices are shared across many triangles through index buffers</li><li>A vertex fetched once can be transformed and reused multiple times</li><li>Connectivity creates strong spatial and temporal locality</li><li>GPU caches are explicitly optimized for this reuse pattern</li></ul><p>As a result:</p><ul><li>small geometry buffers generate large visual detail</li><li>memory traffic stays low relative to rendered pixels</li></ul><p>Gaussian Splats break this assumption:</p><ul><li>Each splat is largely <strong>independent and self-contained</strong></li><li>There is no shared connectivity or index-based reuse</li><li>Neighboring splats may differ in scale, orientation, and access order</li><li>Cache reuse is minimal and short-lived</li></ul><p>The tradeoff is clear:</p><ul><li><strong>Meshes gain efficiency through reuse</strong></li><li><strong>Splats gain flexibility at the cost of memory pressure</strong></li></ul><p>You don’t need smarter math.<br>You need <strong>faster memory</strong>.</p><h3>Why High-Bandwidth Memory Changes Everything</h3><p>This is where NVIDIA’s hardware strategy quietly shines.</p><p>High Bandwidth Memory (HBM):</p><ul><li>sits physically close to the GPU die</li><li>offers massive bandwidth compared to conventional DRAM</li><li>allows far larger scene representations to stay resident</li></ul><p>For Gaussian Splats, this is critical.</p><p>HBM enables:</p><ul><li>more splats per scene</li><li>fewer stalls during traversal</li><li>smoother real-time navigation</li><li>higher-fidelity world representations</li></ul><p>Gaussian Splatting isn’t just about visuals — it’s about <strong>holding a world in memory</strong> and accessing it fast enough to feel real.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yURNYBKTebo_acZV" /></figure><h3>Why Mobile Gaussian Splats Are So Difficult</h3><p>People often ask when Gaussian Splats will “just work” on phones.</p><p>The answer isn’t faster mobile GPUs.</p><p>It’s memory architecture.</p><p>Mobile systems rely on LPDDR:</p><ul><li>lower bandwidth</li><li>higher effective latency</li><li>shared between CPU, GPU, NPU, ISP, and OS</li></ul><p>Gaussian Splats:</p><ul><li>overflow caches almost immediately</li><li>trigger constant memory traffic</li><li>collapse performance under real scenes</li></ul><p>That’s why mobile demos rely on:</p><ul><li>aggressive pruning</li><li>streaming subsets of splats</li><li>heavy compression or simplification</li></ul><p>The bottleneck isn’t rendering quality.<br>It’s <strong>data movement</strong>.</p><h3>From Compute-Centric AI to Memory-Centric AI</h3><p>Gaussian Splatting exposes a deeper shift happening across AI.</p><p>Traditional AI workloads:</p><ul><li>dense tensors</li><li>predictable math</li><li>compute-bound execution</li></ul><p>World models and Physical AI:</p><ul><li>sparse primitives</li><li>massive state</li><li>memory-bound behavior</li></ul><p>This is why ideas like:</p><ul><li>near-memory processing</li><li>in-memory filtering and culling</li><li>processor-in-memory (PIM)</li></ul><p>suddenly make architectural sense.</p><p>Gaussian Splats don’t need more arithmetic.<br>They need <strong>less movement of data</strong>.</p><h3>Rendering Isn’t the Hard Part Anymore</h3><p>The uncomfortable truth is this:</p><blockquote><em>Rendering Gaussian Splats is easy.<br> Supplying them fast enough is not.</em></blockquote><p>That’s why:</p><ul><li>memory architecture matters more than shaders</li><li>bandwidth matters more than peak FLOPs</li><li>system design matters more than clever kernels</li></ul><p>Gaussian Splatting didn’t just change 3D reconstruction.<br>It revealed where modern AI systems actually struggle.</p><p><strong>Not in compute.<br>In memory.</strong></p><h3>Final Thought</h3><p>If you want to understand the future of:</p><ul><li>Gaussian Splats</li><li>Digital Twins</li><li>Robotics simulation</li><li>Physical AI</li></ul><p>stop asking:</p><blockquote><em>“How fast can we render this?”</em></blockquote><p>Start asking:</p><blockquote><strong><em>“How fast can we move the data?”</em></strong></blockquote><p>That’s where the next real breakthroughs will come from.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f7c62e5a0355" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Physical AI Might Replace Manual Labor Faster Than White-Collar Work]]></title>
            <link>https://medium.com/@qsibmini123/physical-ai-might-replace-manual-labor-faster-than-white-collar-work-c89c21d59b9c?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/c89c21d59b9c</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Mon, 08 Dec 2025 15:19:12 GMT</pubDate>
            <atom:updated>2025-12-08T15:19:12.874Z</atom:updated>
            <content:encoded><![CDATA[<h3>Because Robots Can Now Learn From Real-World Digital Twins — Not Just Code</h3><p>Most people assume AI will replace office jobs first — writers, analysts, coders.<br>But physical-AI robots may actually arrive in factories, warehouses, and kitchens <em>faster</em> than AI replaces creative or strategic white-collar roles.</p><p>Why?<br>Because robots no longer need to learn slowly in the real world — they can train inside <strong>Gaussian Splat-based digital twins</strong> captured from real spaces using just a phone.</p><p>What used to take months of risky trial-and-error can now happen overnight.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xz0DIyRaYNl5vMgjDdoeMw.png" /></figure><h3>🧠 Robots No Longer Learn Like Apprentices — They Learn Like Gamers on Fast-Forward</h3><p>In the past, training a robot meant:</p><ul><li>many real mistakes</li><li>hardware breakage</li><li>safety risks</li><li>high cost per attempt</li></ul><p>Simulation changed that — open-world physics, parallel reinforcement learning, infinite trial capacity.</p><p>But <strong>traditional simulation was too slow to model every real room or factory.</strong></p><p>Enter <strong>Gaussian Splats</strong>.</p><ul><li>capture any real environment in minutes</li><li>render it efficiently</li><li>let robots practice inside it millions of times</li><li>transfer those skills back to reality</li></ul><p>Robots don’t just <em>see</em> kitchens or warehouses — they <strong>live inside them digitally first.</strong></p><h3>🌍 Why Gaussian-Splat Worlds Are a Breakthrough for Physical AI</h3><h4>1. They look like the real world</h4><p>Not polygons — real captured texture, lighting, occlusion.</p><h4>2. They’re fast</h4><p>Rendering often bypasses rasterization + ray tracing overhead.<br>Training runs smoothly, even at scale.</p><h4>3. Environment capture is trivial</h4><p>No CAD modeling. No artist pipeline.<br>Just walk with a phone, scan, upload → simulation begins.</p><h4>4. Robots can practice forever</h4><p>Millions of grasp attempts, collisions, drops — with no hardware wear.</p><p>A robot trained in 10,000 virtual kitchens or 500 warehouse layouts <strong>learns generalization</strong>, not memorization.</p><p>In other words:<br><strong>Gaussian splats give robots childhood — at 1000× speed.</strong></p><h3>🏭 Tesla as a Preview of the Future</h3><p>Tesla already uses advanced automation for welding, assembly, and painting.<br>But the company’s next step — <strong>Optimus</strong>, its general-purpose humanoid robot — depends heavily on scalable training.</p><p>With splat-based digital twins of factory floors, Tesla could:</p><ul><li>train robots on line layout before deployment</li><li>test thousands of error cases with zero downtime</li><li>push updated behaviors to every robot simultaneously</li><li>simulate workflow changes overnight</li></ul><p>Training becomes a <strong>software loop</strong>, not a hardware bottleneck.</p><p>This is how manual labor automation accelerates.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8v_aJNPXqS429uY7t29_tQ.png" /></figure><h3>🔥 Why Physical Jobs Might Fall Before Office Work</h3><p>White-collar tasks often require nuance and context.<br> Physical jobs require repetition, consistency, and strength.</p><p>Once robots can learn fast in simulation:</p><ul><li>warehouse picking becomes automatable</li><li>packaging lines no longer need humans</li><li>kitchens, retail stockrooms, hotels begin to shift</li><li>a splat-trained robot learns 1000× faster than an intern</li></ul><p>AI doesn’t need AGI to lift boxes or fold shirts.</p><p>Just good simulation.</p><h3>🤖 Bottlenecks to Solve Before Robots Go Mainstream</h3><p>Physical AI is advancing quickly — but not frictionlessly.</p><p>Modern robots still face challenges:</p><ul><li><strong>Sim-to-real gap</strong> — splats look real, but physics don’t always match.</li><li><strong>Dynamic scenes</strong> — warehouses change, digital twins must update.</li><li><strong>Contact-rich manipulation</strong> — deformable objects (laundry, cables) remain tough.</li><li><strong>Hardware reliability</strong> — simulation is cheap, robots still break.</li><li><strong>Safety</strong> — physical AI must be predictable around humans.</li></ul><p>Simulation gives robots a brain — <br>the body still needs work.</p><h3>Conclusion: The Real Breakthrough Is Acceleration, Not Intelligence</h3><p>We scaled text → we got LLMs.<br>We scaled images → we got diffusion models.<br>Now we are scaling <strong>worlds</strong> — and that is how we get <strong>Physical AI.</strong></p><p>Gaussian Splat digital twins turn the real world into something robots can <em>practice inside</em>, over and over, at machine speed.<br>Not reading instructions. Not watching humans.<br><strong>Learning by doing — millions of times faster than we ever could.</strong></p><p>And the tasks they master first won’t be spreadsheets or slide decks.<br>They’ll be the ones built on repetition, motion, weight, and friction.</p><p>And when that happens — <br><strong>the first jobs AI replaces might involve muscles, not spreadsheets.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c89c21d59b9c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How NVIDIA Is Using 3D Gaussian Splats to Build the Future of Physical AI]]></title>
            <link>https://medium.com/@qsibmini123/how-nvidia-is-using-3d-gaussian-splats-to-build-the-future-of-physical-ai-fad0c7f06c7d?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/fad0c7f06c7d</guid>
            <category><![CDATA[physical-ai]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[neural-rendering]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Mon, 01 Dec 2025 14:20:47 GMT</pubDate>
            <atom:updated>2025-12-01T14:20:47.306Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cnD-glDdHRx5Hq7QNmZBnQ.png" /></figure><p>If the last few years of AI were about teaching models to understand pixels and text, the next era is about something far bigger:<br><strong>teaching AI to understand the physical world.</strong></p><p>At SIGGRAPH 2025, NVIDIA made this shift unmistakably clear. Their keynote didn’t just celebrate graphics breakthroughs — it unveiled a roadmap where <strong>simulation, robotics, and neural rendering converge into a new discipline they call <em>Physical AI</em>.</strong> And at the center of this transformation sits a technology that, until recently, looked like a cool rendering trick:</p><p><strong>3D Gaussian Splats.</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FrFcmv2pXR0w%3Flist%3DPLZHnYvH1qtOaLfQLf1Q8GLpwh8sjPUy2O&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DrFcmv2pXR0w&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FrFcmv2pXR0w%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a1e14739baaee80d6f9a4176362969f3/href">https://medium.com/media/a1e14739baaee80d6f9a4176362969f3/href</a></iframe><p>What started as a fast, beautiful way to reconstruct scenes is now evolving into a foundational representation for machines that <em>see</em>, <em>reason</em>, and <em>act</em> in the real world. Here’s why this matters — and what it means for the future of AI.</p><h3>1. The Big Question: How Do You Teach an AI About Reality?</h3><p>Large language models learned from text. Vision models learned from billions of images. But physical AI — robots that can fold laundry, navigate homes, or collaborate with humans — needs something richer:</p><ul><li><strong>3D structure</strong></li><li><strong>Lighting and reflectance cues</strong></li><li><strong>Temporal change</strong></li><li><strong>Accurate camera behaviors</strong></li><li><strong>Differentiability for simulation</strong></li></ul><p>In short, robots need a <em>world model</em> — a fast, editable, photorealistic representation of reality.</p><p>Traditional 3D formats couldn’t provide this: meshes are brittle, voxels are too heavy, and NeRF-like volumetric models are slow for real-time interaction.</p><p>This is where <strong>Gaussian Splats</strong> come in.</p><h3>2. Gaussian Splats: The World’s Most Efficient Reality Cheat Code</h3><p>A 3D Gaussian Splat is a tiny, soft “blob” in 3D space that holds:</p><ul><li>its position</li><li>its shape (covariance)</li><li>its color</li><li>how much light it absorbs or emits</li></ul><p>Put hundreds of thousands of these together and you get a scene that looks incredibly lifelike — rendered in <strong>real time</strong> on a GPU.</p><p>NVIDIA recognized something profound:<br>Gaussian splats are not just good for <em>view synthesis</em> — they are the first 3D representation that balances:</p><ul><li><strong>fidelity</strong></li><li><strong>speed</strong></li><li><strong>compactness</strong></li><li><strong>differentiability</strong></li><li><strong>real-time rendering</strong></li></ul><p>That makes them perfect for <em>simulation</em> and <em>robotics</em>, not just graphics demos.</p><h3>3. NVIDIA’s Play: Turning Splats Into a Physical-AI Substrate</h3><p>Across SIGGRAPH and their latest research releases, NVIDIA is pushing Gaussian splats deep into their ecosystem:</p><h3>✔ NVIDIA NuRec: Reality → Gaussian Splats</h3><p>NuRec transforms multi-view images, LiDAR, or sensor data into <strong>ray-traced Gaussian Splat scenes</strong>. These can be dropped straight into:</p><ul><li><strong>Omniverse</strong></li><li><strong>Isaac Sim</strong></li><li><strong>digital twin engines</strong></li></ul><p>This makes real-world capture fast, repeatable, and robotics-ready.</p><h3>✔ Cosmos: World Foundation Models</h3><p>Cosmos models learn from — and generate — worlds. They use Gaussian Splat scenes as the geometric backbone for:</p><ul><li>synthetic data generation</li><li>scene understanding</li><li>physical reasoning</li></ul><p>It’s the world model LLMs never had.</p><h3>✔ Physically Embodied Gaussians</h3><p>This is the futuristic piece: a robot keeps a <strong>live simulation</strong> of its environment built from splats + physics particles.</p><p>It renders the splats, compares them to camera views, and corrects its world model in real time.</p><p>It’s basically a robot with an <strong>internal imagination loop</strong> — powered by Gaussian splats.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wz0uLA24jANdLAZUd23ZjQ.png" /></figure><h3>4. 3DGS Just Leveled Up: Meet 3DGUT</h3><p>NVIDIA’s research also introduced <a href="https://research.nvidia.com/labs/toronto-ai/3DGUT/"><strong>3DGUT</strong></a>, an extension to Gaussian Splatting that handles:</p><ul><li>fisheye and wide-angle cameras</li><li>rolling shutter</li><li>real-time reflections and refractions</li><li>arbitrary lens distortions</li></ul><p>This matters because real cameras — especially on mobile devices and robots — are <em>never</em> perfect pinhole lenses.</p><p>3DGUT makes splats usable in messy, real-world settings.<br>It’s the difference between “cool research” and “deployable AI system.”</p><h3>5. The Larger Shift: Simulation-First Robotics</h3><p>NVIDIA’s philosophy is becoming clear:</p><blockquote>Robots should learn in simulation before touching reality.</blockquote><p>And those simulations must be:</p><ul><li>visually accurate</li><li>physically accurate</li><li>easy to update</li><li>differentiable</li><li>compatible with GPU-scale training</li></ul><p>Gaussian splats check every box.</p><p>When splats become the default “world layer,” the entire pipeline becomes unified:</p><ol><li><strong>Capture</strong> your environment → splats</li><li><strong>Simulate</strong> inside Omniverse / Isaac → splats</li><li><strong>Generate synthetic training data</strong> → splats</li><li><strong>Robots imagine &amp; correct mistakes</strong> → splats</li><li><strong>Deploy in reality</strong> with minimal surprise</li></ol><p>This is the first time AI, graphics, and robotics share the same 3D substrate.</p><h3>6. Why This Matters (Even Outside Robotics)</h3><p>This shift affects far more than industrial robots.</p><ul><li><strong>AR/VR</strong>: phones could soon capture splat-based digital twins instantly</li><li><strong>Gaming</strong>: realistic worlds built from real environments</li><li><strong>E-commerce</strong>: real products → splats → interactive viewers</li><li><strong>Digital twins</strong>: cities reconstructed, updated, and simulated with splats</li><li><strong>Content creation</strong>: splat editing tools will become a creative niche</li></ul><p>Gaussian splats are becoming the JPEG of physical spaces — compact, fast, universal.</p><h3>7. The Bottom Line</h3><p>NVIDIA isn’t betting on Gaussian Splats because they look pretty.<br> They’re betting on them because splats solve the hardest part of Physical AI:</p><p><strong>teaching machines what the real world looks like, feels like, and how it behaves.</strong></p><p>With splats as the foundation, NVIDIA is building a pipeline where:</p><ul><li>world capture</li><li>simulation</li><li>synthetic data</li><li>robot learning</li><li>world-model correction</li></ul><p>all use the same representation.</p><p>This is more than a graphics technique.<br>This is the beginning of a new AI stack — one built on <em>world models</em>, not just text or images.</p><p>And Gaussian Splats just became the way machines see the world. With this, robots are learning to be better robots.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fad0c7f06c7d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Future of 3D Gaussian Splats on Mobile Phones]]></title>
            <link>https://medium.com/@qsibmini123/the-future-of-3d-gaussian-splats-on-mobile-phones-5e775bfb1070?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e775bfb1070</guid>
            <category><![CDATA[3d-rendering]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[virtual-reality]]></category>
            <category><![CDATA[mobile-technology]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 07:19:56 GMT</pubDate>
            <atom:updated>2025-11-02T07:19:56.074Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How pocket-sized devices are learning to capture, compress, and render the world in splats.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HS0KbtTDgt1DHv8_Vpj12Q.png" /></figure><h3>1 | The Next Leap After Megapixels</h3><p>Smartphone cameras keep getting sharper, but the real revolution isn’t more pixels — it’s <strong>more dimensions</strong>.</p><p>3D Gaussian Splatting (3DGS) has turned the graphics world upside down since 2023.<br>Instead of polygons or dense neural fields, it paints the world with millions of tiny translucent ellipsoids — “splats” — that softly blend together to form photorealistic 3D scenes.</p><p>The dream?<br>Pull out your phone, circle around an object, and instantly see it as an explorable 3D capture.</p><p>That’s almost here.</p><h3>2 | Why Splats Fit in a Phone’s World</h3><p>Traditional 3D reconstructions depend on heavy geometry or neural networks that choke on mobile GPUs. Gaussian splats are light: they can render directly in screen space without complex ray-marching.</p><p>Each splat stores position, color, size, and shape — enough to reconstruct fine detail while staying memory-efficient.<br>That simplicity makes them perfect for <strong>edge rendering</strong>, where every milliwatt counts.</p><p>Imagine future phones that treat 3D capture like portrait mode: quick, intuitive, and instant.</p><h3>3 | Capturing the World with a Phone Camera</h3><p>Turning handheld video into a clean 3D model sounds easy. It isn’t.</p><ul><li><strong>Pose drift: </strong>tracking can be lost in AR-based platforms; even small errors cause ghosting.</li><li><strong>Lighting variation:</strong> automatic exposure breaks color consistency.</li><li><strong>Rolling shutter:</strong> a tiny shake becomes warped geometry.</li><li><strong>Shallow depth:</strong> phone LiDAR and stereo depth maps are noisy and low-resolution.</li></ul><p>Researchers are already building <strong>self-correcting capture</strong> pipelines that adjust poses as splats train.<br> Future apps will likely guide you live — <em>“too fast,” “move left,” “lighting inconsistent.”</em></p><blockquote><em>A 3DGS-enabled camera app could soon behave like an intelligent director whispering scene coverage advice.</em></blockquote><h3>4 | Rendering Reality, but Gently</h3><p>Desktop GPUs can throw millions of splats per frame. Mobile chips can’t — yet.</p><p>New recent works tackle this problem with clever pruning and level-of-detail hierarchies.<br>Close objects get dense splats; distant ones merge into fewer, larger ellipsoids.<br>This keeps frame rates smooth without tanking quality.</p><p>Hybrid pipelines are also emerging: phones preview scenes instantly, while cloud servers refine and compress the model.<br>Think <em>“Capture on phone, perfect in cloud, view anywhere.”</em></p><h3>5 | What People Will Actually Do with It</h3><p>When 3DGS hits mobile, the uses might explode:</p><ul><li><strong>Virtual try-on:</strong> capture a shoe or bag, preview it on yourself in AR.</li><li><strong>Real estate:</strong> scan a room, share an interactive walkthrough.</li><li><strong>3D social posts:</strong> record “splattable” moments instead of flat videos.</li><li><strong>Robotics and mapping:</strong> small drones or AR navigators capturing real spaces in real time.</li></ul><p>For most users, this won’t feel like scanning — it’ll feel like <em>recording reality itself.</em></p><h3>6 | The Hard Problems Nobody Sees</h3><p>The toughest part of bringing splats to phones might not be GPU performance — it’s data reliability.</p><ul><li><strong>Pose estimation</strong> must be nearly perfect; otherwise splats misalign.</li><li><strong>Dynamic objects</strong> break the assumption of static scenes.</li><li><strong>Lighting</strong> can vary between frames, confusing brightness consistency.</li><li><strong>User UX:</strong> most people won’t walk careful orbits around objects.</li></ul><p>Until capture tools guide users intelligently — highlighting coverage gaps or adjusting parameters on the fly — mobile splats will still trail lab-grade results.</p><h3>7 | Making Splats Lighter and Faster</h3><p>A single high-quality scene can take hundreds of megabytes. That’s where compression research shines.</p><p>Pruning redundant splats, quantizing attributes, and building <strong>progressive LOD streams</strong> cut size by 10× or more.<br>Soon, mobile apps will stream splats like videos — low detail first, full quality if you linger.</p><p>This points toward new infrastructure ideas, making use of clouds. A cloud system hosting 3D splat models, automatically serving detail levels depending on bandwidth and zoom can make application availability possible sooner.<br>It’s the 3D equivalent of YouTube’s adaptive streaming.</p><h3>8 | Looking Ahead (2025 → 2030)</h3><p><strong>2025–26:</strong> consumer apps begin guided splat capture with real-time previews.<br><strong>2027:</strong> hybrid pipelines make full 3D scenes in minutes.<br><strong>2028+:</strong> social platforms allow interactive “3D moments.”<br><strong>Beyond:</strong> mobile SoCs include neural blocks tuned for Gaussian splatting.</p><p>Your phone won’t just <em>take</em> pictures.<br>It’ll <em>remember space.</em></p><h3>9 | The Democratization of 3D</h3><p>Every major shift in imaging — film, digital, HDR — came from putting pro tools into ordinary hands.<br>Gaussian splatting could do that for spatial capture.</p><p>The tech is already fast, compact, and getting friendlier.<br>Give it another hardware cycle or two, and the phrase <em>“Take a photo”</em> might evolve into <em>“Capture a scene.”</em></p><blockquote><em>The next big leap in mobile photography isn’t higher megapixels — it’s higher dimensions.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e775bfb1070" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Standardizing 3D Gaussian Splatting Is So Hard: Datasets, Decoders, and Dollars]]></title>
            <link>https://medium.com/@qsibmini123/why-standardizing-3d-gaussian-splatting-is-so-hard-datasets-decoders-and-dollars-94b172a8732f?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/94b172a8732f</guid>
            <category><![CDATA[compression]]></category>
            <category><![CDATA[standardization]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[mpeg]]></category>
            <category><![CDATA[3d-graphics]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sat, 11 Oct 2025 19:25:04 GMT</pubDate>
            <atom:updated>2025-10-11T19:25:04.145Z</atom:updated>
            <content:encoded><![CDATA[<p>3D Gaussian Splatting has transformed real-time rendering — but turning it into a global standard demands stable datasets, conformance decoders, and industry investment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vnRJY00iJ87n7QkwovacGw.png" /></figure><h3>💥 From a Flash of Brilliance to a Wall of Bureaucracy</h3><p>When <em>3D Gaussian Splatting</em> (3DGS) exploded onto the scene in 2023, the excitement was contagious.<br> Real-time neural rendering of photorealistic 3D scenes — on commodity GPUs — felt like a genuine leap beyond NeRFs.</p><p>Suddenly, everyone from indie developers to major XR labs was playing with splats: glowing ellipsoids that somehow captured reality faster than any volumetric model before.</p><p>But after the demos came the silence.<br>Because after the “wow,” someone always asks:</p><p><em>“So… how do we actually share, store, or stream this stuff?”</em></p><p>That’s when the real work begins.</p><h3>⚙️ The Hard Truth: Standardization Is a Marathon</h3><p>Turning Gaussian splats into a proper <em>interchange format</em> — something that works across devices, platforms, and time — isn’t a matter of better rendering.<br>It’s about <strong>data discipline</strong>, <strong>bitstreams</strong>, and <strong>bureaucracy</strong>.</p><p>You need:</p><ol><li>📦 <strong>Datasets</strong> everyone agrees on</li><li>🧩 <strong>Decoders</strong> that actually match bitstream syntax</li><li>💰 <strong>An economic model</strong> that doesn’t scare off implementers</li></ol><p>Let’s unpack each.</p><h3>1️⃣ Datasets: The Bedrock of Every Codec</h3><p>Every modern MPEG standard begins with <strong>common test conditions</strong> (CTCs): a set of open datasets, metrics, and anchors used to compare coding tools.</p><p>Without them, there’s no reproducibility, no fair benchmarking, and no progress.</p><p>For point clouds, we had <em>8i Voxelized Full Bodies</em> and <em>Ford LiDAR</em>.<br>For immersive video, <em>Technicolor Lightfield</em> and <em>MPEG MIV anchors</em>.<br>For Gaussian splats?<br>Many candidates, but don’t know. There is license issue as well.</p><p>Right now, Gaussian datasets are scattered across labs:</p><ul><li>Some use <strong>Kerbl’s original Mip-NeRF360 captures</strong></li><li>Others reconstruct from <strong>COLMAP or BlendedMVS</strong></li><li>A few use <strong>proprietary dynamic sequences</strong></li></ul><p>A few recent efforts try to impose order — like <strong>GausPCC-1K</strong>, a benchmark dataset proposed in early 2025 for evaluating compression on Gaussian fields.<br>But adoption is still limited.</p><p>Until MPEG’s <em>Ad hoc Group on Gaussian Splat Coding</em> locks in official datasets, no one knows what “good compression” even means.</p><p><em>🧠 </em>Without common data, there is no common ground.</p><h3>2️⃣ Decoders: The Half of Every Codec</h3><p>Every codec has two halves: the encoder (how you compress) and the decoder (how you read).<br>3DGS right now has <em>hundreds</em> of encoders — and almost <strong>no standardized decoder</strong>.</p><p>That’s a problem.</p><p>For standardization, MPEG requires:</p><ul><li><strong>A conformance decoder and bitstreams</strong></li><li><strong>Bit-exact rendering outputs</strong></li><li><strong>A syntax that doesn’t change every six months</strong></li></ul><p>But the splat world is moving too fast:</p><ul><li>Static splats, dynamic splats, deformable splats, hybrid mesh-splats, radiance-encoded splats — each with different fields and matrix shapes.</li><li>Some store <strong>covariance</strong> explicitly; others decompose it into scale + rotation.</li><li>Some do not use AC coefficients</li><li>Some compress with regression; others with quantization.</li></ul><p>A single conformance decoder would need to either:</p><ul><li>Freeze one format and risk obsolescence, or</li><li>Support everything and drown in complexity.</li></ul><p>Neither is fun.<br>So MPEG’s current <strong>Exploration #45 on Gaussian Splat Coding</strong> is wisely keeping it open — identifying <em>what’s stable enough to standardize later.</em></p><p>Until that happens, “decoding” a Gaussian Splat file is still more art than science.</p><h3>3️⃣ Dollars: The Elephant in the Lab</h3><p>Even when the science stabilizes, the business has to make sense.</p><p>Standardization isn’t free — it costs:</p><ul><li>💼 <strong>Patent licensing</strong> (MPEG LA or AOM-like pools)</li><li>💿 <strong>Software maintenance</strong> (reference implementations, compliance tools)</li><li>🌐 <strong>Ecosystem buy-in</strong> (device vendors, streaming services, content creators)</li></ul><p>The risk?<br>If Gaussian splat coding becomes an <strong>IP-heavy</strong> standard, adoption could die before it begins. Developers already burned by MPEG PCC’s patent landscape will hesitate to build another closed ecosystem.</p><p>That’s why some in the community are quietly advocating for an <strong>open baseline profile</strong>, possibly royalty-free, similar to <em>AOM’s AV1</em> model.</p><p>But until big vendors (Apple, Meta, Qualcomm, NVIDIA) show alignment, standardization committees won’t move quickly.</p><p>And big vendors won’t align until there’s clear commercial value — AR streaming, holographic capture, volumetric chat, something real.</p><p>Which brings us to the biggest bottleneck of all.</p><h3>🧩 The Real Problem: No Killer App (Yet)</h3><p>Unlike point clouds (used in automotive LiDAR and 3D maps) or immersive video (used in XR concerts), Gaussian splats don’t yet have a <strong>flagship deployment scenario</strong>.</p><p>They <em>could</em> power:</p><ul><li>XR environments (low-latency free-viewpoint scenes)</li><li>Web-native 3D previews</li><li>Volumetric telepresence</li><li>Neural asset libraries for game engines</li></ul><p>…but these are still experiments, not billion-dollar pipelines.</p><p>And without a clear commercial driver, <strong>funding the slog</strong> — test data curation, encoder-decoder validation, committee meetings — becomes difficult.</p><h3>🧠 Where MPEG Fits In (Quietly)</h3><p>MPEG isn’t ignoring splats.<br>The <em>Ad hoc Group on Gaussian Splat Coding</em> launched in 2025 is explicitly exploring:</p><ul><li>Datasets</li><li>Test conditions</li><li>Evaluation metrics</li><li>Potential carriage formats (possibly extending V3C or OMAF)</li></ul><p>It’s a cautious but concrete start.<br>In MPEG terms, an “Exploration” can evolve into a <strong>Call for Proposals</strong> within 12–18 months — if the community rallies around stable data and codecs.</p><h3>🔍 The Balancing Act Ahead</h3><p>Gaussian splatting is still at that awkward adolescence between research and standard.</p><p>The community faces three hard choices:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5cwXvWxqv3xz2k95_IHOXA.png" /></figure><p>No choice is wrong — but indecision will delay standardization indefinitely.</p><h3>🎯 Takeaway</h3><p>3D Gaussian splatting changed how we think about neural rendering — fast, expressive, and artist-friendly.<br>But turning it into a <em>standard</em> means dealing with far less glamorous work: collecting datasets, writing decoders, and sorting out dollars.</p><p>If the community can stay patient — and pragmatic — then Gaussian splats might just become the backbone of next-gen immersive media.</p><p>But for now, it’s not about speed.<br>It’s about <strong>structure</strong>.</p><h3>🧭 Conclusion: From Demos to Deployability</h3><p>Gaussian splatting gave us the first <em>truly real-time</em> neural 3D representation.<br>But demos alone don’t make an ecosystem.</p><p>To become the “JPEG of 3D scenes,” 3DGS needs:</p><ol><li>📦 <strong>Datasets</strong> everyone agrees on</li><li>🧩 <strong>Decoders</strong> that actually match bitstream syntax</li><li>💰 <strong>An economic model</strong> that doesn’t scare off implementers</li></ol><p>As with every standardization journey, this one won’t be fast.<br>But it’s already begun — and for 3D Gaussian splatting, that’s the most important splat of all.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94b172a8732f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Pixar to TikTok: Why Gaussian Splats Could Be the Next Big Thing in 3D]]></title>
            <link>https://medium.com/@qsibmini123/from-pixar-to-tiktok-why-gaussian-splats-could-be-the-next-big-thing-in-3d-0ee90f0c4ca5?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ee90f0c4ca5</guid>
            <category><![CDATA[sns]]></category>
            <category><![CDATA[3d-modeling]]></category>
            <category><![CDATA[vr]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[compression]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 21 Sep 2025 02:46:58 GMT</pubDate>
            <atom:updated>2025-09-21T02:46:58.105Z</atom:updated>
            <content:encoded><![CDATA[<h3>🎬 From Polygons to Pixels to Splats</h3><p>Pixar changed the world of movies by perfecting <strong>polygon-based graphics</strong> — every 3D object you see in Toy Story or Finding Nemo is made of tiny triangles stitched together.</p><p>Fast-forward to TikTok and YouTube Shorts: billions of people now consume <strong>bite-sized visual content</strong> daily. What’s next? A shift from flat videos to <strong>immersive 3D clips</strong> that you can step inside.</p><p>The math powering this shift comes from an unlikely hero: <strong>Gaussian Splats</strong>.</p><h3>🌌 What Exactly Are Gaussian Splats?</h3><p>Imagine painting a 3D scene not with solid blocks, but with <strong>millions of paint blobs</strong>. Each blob is fuzzy, colorful, and slightly transparent — like tiny floating ellipsoids. Put enough of them together, and you get a photorealistic 3D world.</p><p>This method is:</p><ul><li><strong>Fast</strong>: rendering happens in real time, even on modest GPUs.</li><li><strong>Smooth</strong>: no jagged polygons, no blocky artifacts.</li><li><strong>Flexible</strong>: works beautifully with messy real-world data like photos or scans.</li></ul><p>👉 In short: splats are like the <strong>pixels of the 3D future</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QvCdorDNODu3ZzUN" /><figcaption>Polygons vs Splats: two ways of building a 3D world</figcaption></figure><h3>🚀 Why Splats Matter in 2025</h3><p>Here’s where Gaussian Splats could change the game:</p><p><strong>Movies &amp; Animation</strong></p><ul><li>Studios can create <strong>faster, cheaper previews</strong> of complex scenes.</li><li>Realistic lighting and soft edges come “for free.”</li></ul><p><strong>Gaming &amp; VR</strong></p><ul><li>Games built on splats could run smoother without needing ultra-high-end graphics cards.</li><li>Virtual reality scenes become lighter to render — think more fluid, immersive VRChat or metaverse spaces.</li></ul><p><strong>Social Media (TikTok, IG, YouTube)</strong></p><ul><li>Creators could scan themselves or objects in 3D and post them instantly.</li><li>Imagine scrolling TikTok and being able to “step inside” a dance video.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qpVFFD70HbRhbeGr" /><figcaption>From flat videos to interactive 3D splats on your social feed</figcaption></figure><h3>⚠️ The Catch: Not All Smooth Sailing</h3><p>Of course, Gaussian Splats aren’t perfect yet.</p><ul><li><strong>Data Size</strong>: storing millions of blobs is heavy. Compression research is key.</li><li><strong>Editing Tools</strong>: artists are used to polygons; splats need new workflows.</li><li><strong>Hardware Limits</strong>: not every phone or headset can handle real-time splats… yet.</li></ul><p>But these are <strong>solvable problems</strong> — and they’re being tackled right now by researchers and open-source communities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*c-tYL5iAG674CuaJ" /><figcaption>Challenges ahead: storage, editing, and performance</figcaption></figure><h3>🔮 The Future: Splats Everywhere</h3><p>So, will splats actually replace polygons? Probably not overnight. But in certain areas — <strong>photorealistic captures, VR, AR, social media content</strong> — they’re already proving superior.</p><p>Think about it: the same math that could speed up Pixar’s next blockbuster might also power the next viral TikTok trend.</p><p>The jump from <strong>Pixar → TikTok</strong> shows just how universal splats might become.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nRL7Qnw-iWhqZfCaIRpmZQ.png" /><figcaption>From Pixar to TikTok: splats bridging high-end film and everyday content</figcaption></figure><h3>✨ Final Thought</h3><p>The internet started with text, then shifted to images, then video. <strong>3D splatted worlds might be the next big shift.</strong></p><p>Maybe the next viral TikTok won’t be a flat clip at all, but a <strong>Gaussian splat world you can walk through.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ee90f0c4ca5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gaussian Splats Are Becoming the JPEG of 3D: Why 2025 Is the Breakout Year]]></title>
            <link>https://medium.com/@qsibmini123/gaussian-splats-are-becoming-the-jpeg-of-3d-why-2025-is-the-breakout-year-ac841ed39440?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac841ed39440</guid>
            <category><![CDATA[immersive-media]]></category>
            <category><![CDATA[3d-rendering]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[3dgs]]></category>
            <category><![CDATA[gltf]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 07 Sep 2025 14:21:42 GMT</pubDate>
            <atom:updated>2025-09-07T14:21:42.804Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why this “JPEG of 3D” is jumping from labs to phones, game engines, and standards</h3><p><strong>TL;DR:</strong> Gaussian splats moved from research curiosity to production reality in 18 months. We now have <strong>real-time 3D on consumer devices</strong>, <strong>90% smaller files</strong>, <strong>city-scale mobile streaming</strong>, and even <strong>early standardization in glTF</strong> — all while the core methods keep getting faster and more geometrically faithful.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lGD9Kahoeka0dyAq" /></figure><h3>1) A two-minute primer (and why splats beat meshes — for now)</h3><p>3D Gaussian Splatting (3DGS) represents a scene as millions of anisotropic “blobs,” each with position, color, opacity, and orientation. Because these primitives <strong>rasterize extremely fast</strong>, you can hit <strong>real-time novel-view synthesis</strong> without baking or heavy neural inference — something classic NeRFs struggled to do at 1080p. The seminal paper demonstrated <strong>≥30 FPS real-time rendering</strong> with strong fidelity.</p><h3>2) Receipts: adoption hit escape velocity in 2024–2025</h3><ul><li><strong>Niantic Scaniverse</strong> shipped on-device 3DGS capture and later <strong>open-sourced the .SPZ format</strong>, citing <strong>~90% smaller</strong> files vs. PLY (e.g., 25 MB vs. 250 MB).</li><li><strong>Meta’s Hyperscape</strong> showcased photorealistic Gaussian-splat scenes on Quest as a public demo of where consumer VR is headed.</li><li>Mainstream tech press is framing splats as a potential <strong>“JPEG moment for spatial computing,”</strong> capturing the jump in fidelity and speed.</li></ul><h3>3) The big news: standardization is underway</h3><p>August 2025 brought a watershed: <strong>Khronos + OGC announced adding 3D Gaussian Splats to the glTF ecosystem</strong>, with <strong>SPZ</strong> as a compact container. This is the difference between clever demos and an <strong>interoperable supply chain</strong> across tools, engines, and the web.</p><p>The standards community is also convening <strong>SIGGRAPH workshops</strong> to steer formats, compression, and best practices — clear momentum toward a stable interchange.</p><h3>4) Performance keeps leaping: from desktops to phones (and cities)</h3><ul><li><strong>FlashGS (CVPR 2025)</strong>: reworked the rasterization pipeline to keep <strong>4K, large-scale scenes</strong> in real-time — think city blocks, not coffee mugs.</li><li><strong>Voyager (Jun 2025)</strong>: a cloud-client design that streams <strong>only newly visible splats</strong>, showing <strong>&gt;100× less data</strong> and up to <strong>8.9× client speedups</strong> for <strong>city-scale</strong> mobile experiences.</li><li><strong>LiveGS (2025)</strong> pushes toward <strong>live free-viewpoint video</strong> with a splat-centric broadcast pipeline, pointing at volumetric streaming beyond prerecorded scans.</li></ul><h3>5) Storage &amp; transmission: from “too big” to “practical”</h3><p>A wave of research is shrinking splats <strong>without</strong> wrecking quality:</p><ul><li><strong>SPZ</strong> (open-sourced by Niantic): lightweight, <strong>~90% smaller</strong> than PLY in practice.</li><li><strong>HAC++</strong>: chasing <strong>up to 100× compression</strong> via learned factorization and hierarchy.</li><li><strong>Surveys (2024–2025)</strong> catalog dozens of <strong>3DGS compression</strong> and compaction methods that make mobile/standalone VR far more realistic.</li><li><strong>Fast, feed-forward compression</strong> (FCGS) cuts <em>minutes</em> of optimization to <strong>seconds</strong> per scene — critical for production pipelines.</li></ul><p>MPEG even opened an <strong>Explorations</strong> track on <strong>Gaussian Splat Coding</strong> (use cases &amp; requirements), a strong signal that formal coding standards may follow.</p><h3>6) Beyond “pretty pictures”: robotics, SLAM, and mapping</h3><ul><li><strong>SplaTAM</strong> and <strong>GS-SLAM</strong> adapt splats to <strong>online mapping &amp; tracking</strong>, bringing fast rendering + dense optimization into robotics.</li></ul><p>These lines of work nudge splats from <strong>static capture</strong> toward <strong>persistent, updatable world models</strong> — relevant to AR cloud maps, autonomous navigation, and digital twins.</p><h3>7) Tooling you can actually use (today)</h3><ul><li><strong>gsplat</strong> (Nerfstudio): a CUDA rasterizer and Python API that’s <strong>faster &amp; more memory-efficient</strong> than early code; active releases through 2025.</li><li><strong>Nerfstudio’s Splatfacto</strong> family: battle-tested training &amp; eval pipelines, plus <strong>Splatfacto-W</strong> for unconstrained image collections.</li><li><strong>Official 3DGS repo</strong> (GraphDECO): reference implementation and pretrained models — good ground truth for experiments.</li></ul><h3>8) What’s next (and where the gaps still are)</h3><p><strong>Interoperability &amp; authoring.</strong> glTF integration + SPZ makes exchange simpler, but creators still need <strong>DCC-grade editing</strong> (selections, boolean ops, lighting semantics) and <strong>mixed pipelines</strong> (splat ↔ mesh ↔ point cloud). Early signs show engines and DCC tools adding native support, but workflows aren’t “one-click” yet.</p><p><strong>Mobile &amp; standalone VR.</strong> Research like <strong>Voyager</strong> and <strong>FlashGS</strong> shows the path to <strong>city-scale, 4K, low-latency</strong> experiences; the next cut is <strong>robust LOD + streaming standards</strong> so apps don’t reinvent the wheel.</p><p><strong>Compression &amp; rights.</strong> We’ll see <strong>standardized bitstreams</strong> (think “AV1-for-splats”), plus asset-protection and provenance baked into formats as splats spread through games, films, and UGC. <strong>MPEG Explorations</strong> are the canary here.</p><p><strong>Dynamic &amp; editable worlds.</strong> 4D-GS and <strong>LiveGS</strong> hint at <strong>streaming, relightable</strong> worlds; add <strong>2DGS-grade geometry</strong> and you’re close to production-quality <strong>interactive volumetric scenes</strong> — with artist-friendly controls.</p><h3>9) If you only remember five numbers…</h3><ul><li><strong>≥30 FPS</strong> real-time rendering from the original 3DGS paper.</li><li><strong>~90% smaller</strong> files via <strong>SPZ</strong> (e.g., 25 MB vs. 250 MB).</li><li><strong>&gt;100× less data</strong> and <strong>up to 8.9× speedup</strong> for city-scale mobile streaming (Voyager).</li><li><strong>4K real-time</strong> on large scenes with <strong>FlashGS</strong>.</li><li><strong>glTF + 3DGS</strong> integration announced by <strong>Khronos/OGC</strong> in <strong>Aug 2025</strong>.</li></ul><h3>10) Bottom line</h3><p>2023 proved splats could be <strong>fast and gorgeous</strong>.<br>2024 proved they could be <strong>geometrically sound</strong> and <strong>mobile-friendly</strong>.<br><strong>2025 is proving they can be <em>standardized</em> and <em>scaled</em>.</strong></p><p>If you’re building XR, robotics, mapping, or volumetric video, <strong>Gaussian splats are no longer an experiment — they’re a platform</strong>. The next twelve months will be about <strong>workflow polish, streaming standards, and creator tools</strong> catching up with the science.</p><h3>Credits &amp; Further Reading</h3><ul><li><a href="https://arxiv.org/abs/2308.04079">3DGS</a> (Kerbl et al., SIGGRAPH 2023) and project page.</li><li><a href="https://dl.acm.org/doi/10.1145/3641519.3657428">2DGS</a> (SIGGRAPH 2024) for geometry-accurate splats.</li><li><a href="https://arxiv.org/abs/2310.08528">4D-GS</a> for dynamic scenes.</li><li><a href="https://openaccess.thecvf.com/content/CVPR2025/papers/Feng_FlashGS_Efficient_3D_Gaussian_Splatting_for_Large-scale_and_High-resolution_Rendering_CVPR_2025_paper.pdf">FlashGS</a> (CVPR 2025), <a href="https://arxiv.org/html/2506.02774v2">Voyager</a> (2025), <a href="https://dl.acm.org/doi/10.1145/3721257.3734033">LiveGS</a> (2025).</li><li><a href="https://scaniverse.com/news/spz-gaussian-splat-open-source-file-format">SPZ</a>: open-source format and size numbers.</li><li>Standards: Khronos + OGC glTF extensions; SIGGRAPH workshop slides.</li></ul><p><strong>Follow me for more on compression, XR, and 3D standards — I cover the bleeding edge where research turns into production!</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac841ed39440" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Hidden Cost of Real-Time 3D: Why Bandwidth is the Next Big Bottleneck]]></title>
            <link>https://medium.com/@qsibmini123/the-hidden-cost-of-real-time-3d-why-bandwidth-is-the-next-big-bottleneck-9e1687d2b810?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/9e1687d2b810</guid>
            <category><![CDATA[point-cloud-compression]]></category>
            <category><![CDATA[point-cloud]]></category>
            <category><![CDATA[volumetric-media]]></category>
            <category><![CDATA[bandwidth-limit]]></category>
            <category><![CDATA[immersive-media]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 31 Aug 2025 14:04:39 GMT</pubDate>
            <atom:updated>2025-08-31T14:04:39.670Z</atom:updated>
            <content:encoded><![CDATA[<p>As immersive technologies like VR, AR, and digital twins move closer to mainstream adoption, a quiet bottleneck is already forming, other than headset resolution or GPU power. It’s bandwidth.</p><p>In the rush to build lifelike virtual worlds, we’ve overlooked a fundamental constraint: how much data can realistically move through the network in real time. And as experiences shift from flat video to point clouds, volumetric video, and Gaussian splats, the gap between what we can create and what we can deliver is only widening.</p><p><em>Immersive media like volumetric video and digital twins face bandwidth bottlenecks when streaming in real time.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rrsv8llc-QnMZqgV_gighQ.png" /></figure><h3>Why Bandwidth Matters More Than Ever</h3><p>Traditional video streaming already consumes massive bandwidth, but immersive media multiplies the problem. Instead of sending a single camera view, we’re transmitting millions of points, surfaces, and textures in 3D.</p><ul><li><strong>Volumetric video</strong> can demand <strong>gigabits per second</strong> of raw data.</li><li><strong>Cloud-rendered VR</strong> adds not just video frames, but constant motion prediction.</li><li><strong>Digital twins</strong> in manufacturing or urban planning require persistent, real-time updates.</li></ul><p>Even with 5G and fiber, these demands exceed what most networks can handle reliably. And unlike Netflix buffering, latency in VR isn’t just annoying — it can break immersion or cause motion sickness.</p><h3>Compression Beyond Size Reduction</h3><p>For decades, compression was about shrinking file sizes. With immersive media, it’s about survival. The key question isn’t <em>how small can the data get</em>, but <em>how fast can it stream without breaking the illusion?</em></p><ul><li><strong>Point Cloud Compression (MPEG PCC)</strong>: Transforming millions of 3D points into manageable bitstreams.</li><li><strong>AI-driven codecs</strong>: Learning patterns in motion and geometry to predict and reduce redundancy.</li><li><strong>Gaussian splatting</strong>: Representing reality as compact 3D ellipsoids that stream more efficiently than raw meshes.</li></ul><p>Each technique balances size, quality, and latency. But none fully solve the challenge alone.</p><h3>Edge Computing: A Partial Lifeline</h3><p>One promising approach is moving computation closer to the user — at the network edge. Instead of streaming raw volumetric video, servers could pre-compress, filter, or adaptively stream depending on user position.</p><p>Think of it like “foveated streaming” for bandwidth: you don’t need the entire 3D scene in ultra-high resolution if you’re only looking at a small portion. Edge nodes can prioritize where the bandwidth goes.</p><p>But this raises new questions: Who controls the edge? Telecoms? Cloud providers? Startups? Whoever solves this becomes the new gatekeeper of immersive media.</p><h3>The Future: Bandwidth as a Competitive Advantage</h3><p>It’s tempting to think compression is a purely technical problem, but it’s increasingly geopolitical and economic. Countries with robust network infrastructure may lead immersive adoption simply because they can deliver it. Platforms with proprietary codecs or edge networks will gain leverage over content creators.</p><p>The story of the next decade of immersive media may not be about hardware or algorithms alone — but about who controls the invisible pipes that carry our worlds.</p><h3>Closing Thoughts</h3><p>Immersive media promises breathtaking possibilities, but without solving bandwidth, the dream remains out of reach. Compression, smarter representations like Gaussian splats, and edge computing are part of the answer. Yet ultimately, bandwidth itself may become the true frontier — the hidden cost that decides who experiences the future first.</p><p>👉 What do you think: will compression innovations be enough, or will bandwidth always be the limiting factor?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9e1687d2b810" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Roadblocks to Mass VR Adoption]]></title>
            <link>https://medium.com/@qsibmini123/the-roadblocks-to-mass-vr-adoption-6b5a9c68cee1?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b5a9c68cee1</guid>
            <category><![CDATA[vr]]></category>
            <category><![CDATA[head-mount-display]]></category>
            <category><![CDATA[metaverse]]></category>
            <category><![CDATA[immersive-media]]></category>
            <category><![CDATA[virtual-reality]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 24 Aug 2025 16:59:54 GMT</pubDate>
            <atom:updated>2025-08-24T16:59:54.416Z</atom:updated>
            <content:encoded><![CDATA[<p>Virtual reality (VR) has been “the next big thing” for nearly a decade. We’ve seen flashy headset launches, metaverse promises, and jaw-dropping demos. And yet, despite billions in investment, VR has not yet gone truly mainstream.</p><p>So what’s standing in the way? Here are the biggest roadblocks between VR today and the future where it’s as common as smartphones.</p><h3>1. The Price Barrier</h3><p>High-end VR headsets still cost anywhere from $400 to over $1,000. Add a powerful PC or console to run the most immersive experiences, and the price skyrockets. While standalone headsets like Meta Quest have made progress, the cost is still too steep for most casual users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p14WeM7S903IX-_FTul7Rw.png" /></figure><p>Mainstream adoption usually requires devices to feel like a no-brainer purchase — like smartphones did once carriers subsidized them. VR hasn’t hit that sweet spot yet.</p><h3>2. Comfort and Usability</h3><p>Even the best modern headsets are bulky, heavy, and uncomfortable to wear for long periods. Motion sickness remains a real issue, especially for first-time users. Battery life is short, and fiddling with straps, controllers, and tracking systems can be a hassle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YUFEeWzAPVetXqhqxciONQ.png" /></figure><p>For VR to become daily tech, it needs to feel as seamless and natural as slipping on a pair of glasses. We’re not there yet.</p><h3>3. The Content Drought</h3><p>Ask someone why they bought a console and they’ll point to exclusive games. Ask someone why they should buy a VR headset, and the answer isn’t always clear.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cCADrwFtbYW_5Izs8MQwhg.png" /></figure><p>Yes, there are brilliant VR titles like <em>Half-Life: Alyx</em> and addictive fitness apps like <em>Beat Saber</em>. But there’s no single “killer app” that convinces the average consumer they <em>need</em> VR. Until there’s must-have content, adoption will stay niche.</p><h3>4. Social Isolation vs. Social Potential</h3><p>VR promises connected virtual worlds, but most users experience it as an isolated activity — headset on, cut off from the real room. That creates friction when using it at home around family or friends.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n7XZQrmDSn5Rb9WcbUn_XA.png" /></figure><p>Ironically, the tech that promises the <em>metaverse</em> often makes people feel less social in the real world. Finding ways to integrate VR naturally into social life will be key.</p><h3>5. Fragmentation of Platforms</h3><p>Unlike smartphones (where iOS and Android dominate), VR is scattered across Meta, Sony, Valve, HTC, Apple, and more. Each has its own hardware, software store, and standards.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CgumcTZpO80KULvderaTTQ.png" /></figure><p>This makes life harder for developers and confusing for consumers. Until there’s more unification — or at least cross-compatibility — VR will struggle to grow beyond fragmented communities.</p><h3>6. The Reality Check of Daily Life</h3><p>Finally, VR has yet to prove its <em>daily utility</em>. Smartphones became universal because they weren’t just for fun — they handled email, navigation, payments, photography, and more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*10R5sOV5feJ69sAq_RI37Q.png" /></figure><p>VR, so far, is still framed as entertainment. To move into the mainstream, it needs to find those everyday killer use cases — whether that’s work, education, healthcare, or social connection.</p><h3>The Bottom Line</h3><p>Virtual reality isn’t going away. The hardware keeps improving, companies like Apple and Meta are doubling down, and new applications are being explored every year.</p><p>But adoption won’t skyrocket until VR solves its core roadblocks: cost, comfort, content, and convenience. Once those barriers fall, VR might finally make the leap from hype to habit.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b5a9c68cee1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gaussian Splatting: A Glimpse Into the Future of Immersive Media]]></title>
            <link>https://medium.com/@qsibmini123/gaussian-splatting-a-glimpse-into-the-future-of-immersive-media-8ddd234cad2f?source=rss-7f85e88692f0------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ddd234cad2f</guid>
            <category><![CDATA[3d-graphics]]></category>
            <category><![CDATA[neural-radiance-fields]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[graphics]]></category>
            <category><![CDATA[immersive-media]]></category>
            <dc:creator><![CDATA[Mario Lee]]></dc:creator>
            <pubDate>Sun, 17 Aug 2025 13:28:50 GMT</pubDate>
            <atom:updated>2025-08-17T13:28:50.318Z</atom:updated>
            <content:encoded><![CDATA[<p>Immersive media is rapidly evolving, and new ways of representing 3D worlds are changing how we think about virtual reality, augmented reality, and digital content creation. Among the most exciting breakthroughs in recent years is <strong>Gaussian Splatting</strong> — a technique that offers both <strong>photorealistic quality</strong> and <strong>real-time rendering speeds</strong>.</p><p>This post introduces Gaussian Splatting, explains why it’s gaining attention, and looks ahead at its potential applications and challenges. For your attention, I have made a simple Gaussian Splat below of a teddy bear I have knitted, using an App called <a href="https://scaniverse.com/">Scaniverse</a>. I had to put it in a really low resolution to upload in Gif format, but the actual output is really nice.</p><figure><img alt="Gaussian Splat of my teddy bear" src="https://cdn-images-1.medium.com/max/320/1*DN0_rB4eL5-Y1cjTT0JhRQ.gif" /><figcaption>Gaussian Splats of my Teddy Bear</figcaption></figure><h3>What is Gaussian Splatting?</h3><p>Gaussian Splatting is a novel method for representing 3D scenes, first presented in this <a href="https://arxiv.org/pdf/2308.04079">paper</a>. Instead of describing geometry with meshes (triangles) or point clouds (discrete points), it uses <strong>3D Gaussian Splats</strong> — smooth splats that carry information about position, spherical harmonics(color), opacity, scale, and rotation. Additional attributes can also be present depending on which Gaussian Splatting technology is used, which is advancing very fast.</p><p>The Gaussians are combined, then they “splat” onto the screen, forming highly realistic images that can be viewed from any angle. The key advantage is that the rendering happens <strong>directly and efficiently</strong>, without the need for heavy neural network computations like in Neural Radiance Fields (NeRFs). Additionally, the time needed for training is reduced significantly compared to conventional neural radiance fields.</p><p>The result? A way to <strong>capture, store, and render 3D scenes in real-time</strong> with a striking balance of speed and quality. For details, please refer to the paper cited above, or the official youtube video of INRIA below.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FT_kXY43VZnk%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DT_kXY43VZnk&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FT_kXY43VZnk%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/6c11b5535cc2d1bc973a3bc45f3825ba/href">https://medium.com/media/6c11b5535cc2d1bc973a3bc45f3825ba/href</a></iframe><h3>Why It Matters</h3><p>For years, immersive media has faced a tradeoff: either you get high-quality 3D scenes (but with slow rendering and heavy storage needs), or you get fast, lightweight experiences (but with reduced realism).</p><p>Gaussian Splatting is breaking this tradeoff by offering:</p><ul><li><strong>Photorealism</strong> — Detailed lighting, textures, and geometry.</li><li><strong>Real-time Rendering</strong> — Interactive framerates, even on consumer GPUs.</li><li><strong>Scalability</strong> — Works for small objects or entire large environments.</li></ul><p>This balance makes it a promising foundation for the <strong>next generation of 3D content</strong>.</p><h3>Potential Applications</h3><ol><li><strong>Virtual and Augmented Reality (VR/AR)</strong><br> Realistic 3D environments can be streamed and interacted with, making VR worlds more convincing and AR overlays more natural.</li><li><strong>Volumetric Video</strong><br> Performances, concerts, or sports events could be captured and replayed as immersive 3D experiences with lifelike presence.</li><li><strong>Digital Twins and Simulation</strong><br> Real-world spaces (factories, cities, landmarks) can be scanned and reproduced in high fidelity for design, training, or education.</li><li><strong>Gaming and Metaverse</strong><br> Game developers could build vast, richly detailed worlds without the same bottlenecks of traditional graphics pipelines.</li><li><strong>Many other!!!</strong></li></ol><h3>Challenges and Open Problems</h3><p>While promising, Gaussian Splatting is not a solved problem. Researchers and developers are actively tackling key challenges:</p><ul><li><strong>Data Size</strong>: Storing millions of Gaussians is still heavy — compression techniques are needed.</li><li><strong>Editing and Interaction</strong>: Modifying Gaussian-based scenes is harder than editing meshes or point clouds.</li><li><strong>Standardization</strong>: Current pipelines in gaming engines and standards bodies lack built-in support.</li><li><strong>Hardware Optimization</strong>: While rendering is efficient, making it mobile-friendly for AR glasses remains a challenge.</li></ul><h3>The Road Ahead</h3><p>Gaussian Splatting is still young, but its rapid adoption in both research and open-source communities shows how quickly it could move into mainstream use.</p><ul><li>For <strong>developers</strong>, it promises a way to create richer worlds with fewer compromises.</li><li>For <strong>researchers</strong>, it opens new doors in graphics, compression, and real-time rendering.</li><li>For <strong>everyday users</strong>, it could mean more immersive games, lifelike AR experiences, and new forms of media yet to be imagined.</li></ul><p>In short, Gaussian Splatting isn’t just another graphics technique. It’s a glimpse into a <strong>future where digital and physical realities blend seamlessly</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ddd234cad2f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>