ComfyAI·Aug 15CLIP’s Image Embeddings Were Never “Uncalibrated”Over the past couple of years, a story has quietly become common knowledge: image-text contrastive learning like CLIP only pulls images and…
ComfyAI·Aug 13Solving Geometry by Drawing the AnswerThere’s a CVPR 2026 paper that uses a diffusion model to solve geometry problems, and the part I find interesting is what it doesn’t do…
ComfyAI·Aug 12Is RL Actually Teaching Multimodal Models to See, or Just Sharpening the Language Model Underneath?Everyone is doing RL post-training on multimodal LLMs right now to push up visual reasoning scores. The benchmark numbers go up, the papers…
ComfyAI·Aug 11Attention Alone Is Weak: Selection-as-Nonlinearity and the Budget Problem Hiding Inside SoftmaxIn theory, a self-attention block with independent input and output projections is a universal approximator — it can fit anything. But…
ComfyAI·Aug 10SNR–t Bias: Why Diffusion Samples Drift Off Their Own TimelineWhen you train a diffusion model, the signal-to-noise ratio of every sample at every timestep t is pinned down by the sigma schedule. The…
ComfyAI·Aug 9Why Diffusion Models Don’t Just Memorize (And What Happens When They Do)Diffusion models memorize. Sometimes the image they generate is nearly pixel-identical to a specific training image.
ComfyAI·Aug 8Your Diffusion Model Already Knows Which Images Are Going to Be GoodSame prompt, swap the random seed, and the quality of what a diffusion model hands you swings all over the place. So you reroll. And reroll…
ComfyAI·Aug 7Blurry by Design: What a Newborn’s Bad Eyesight Can Teach Self-Supervised LearningNewborn vision is blurry, nearly colorless, and changes continuously over time. Only across the first year does it gradually sharpen and…
ComfyAI·Aug 6The Information Horizon: Why Clever Visual Token Pruning Stops Beating RandomA VLM chops a single image into hundreds of visual tokens. Attention is quadratic, image resolutions keep climbing, and inference gets…
ComfyAI·Aug 5Dataset Distillation’s Gains Are Coming From the Labels, Not the ImagesDataset distillation (DD) compresses something like ImageNet down to a few dozen synthetic images per class — images produced by treating…
