Researchers Comment on Text-to-Image Diffusion Scaling Laws
Conversation questions use of classifier-free guidance after praise for a scaling study.
Anthropic research engineer @gallabytes called a scaling study on text-to-image diffusion models the rare good ML paper of 2026. Prime Intellect researcher @kalomaze replied that the subfield adopts classifier-free guidance without checking why conditional samples fall off manifold without extrapolation at inference time. @cloneofsimo noted autoregressive image models and some language diffusion models also use it. @kalomaze answered that the practice remains a poorly understood heuristic regardless of model family.
Combined views
18.6K
9 posts, first seen 5d ago
Reactions from ranked influencers
9 posts
love to see basic science applied to image gen the rare "good ml paper in 2026"
🧵Scaling Laws for T2I Diffusion We do a scaling laws ladder for t2i diffusion models all the way from 60M to 2B, spanning 3 orders of magnitude in training flops
@gallabytes rare for the kind of subfield that likes to adopt methods such as "classifier free guidance" without asking why the conditional samples aren't naturally on manifold without inference time extrapolation lmao
@kalomaze @gallabytes by subfield do you mean diffusion models ? there are good science in diffusion models
Pretty common misconception, as in: 1) Autoregressive models for image domain also need CFGs [1,2] 2) For languge domains, diffusion models dont need CFGs [3] https://arxiv.org/abs/2410.13848 https://arxiv.org/abs/2608.00146
@gallabytes rare for the kind of subfield that likes to adopt methods such as "classifier free guidance" without asking why the conditional samples aren't naturally on manifold without inference time extrapolation lmao
@cloneofsimo i mean, i dont see how this is a misconception as much as it is a statement about how the field seems to struggle with faithfully modeling high dimensions in any regime
@cloneofsimo the ideal case is "you should not have to force the network to extrapolate beyond what conditioning supplies", and whether or not that's an artifact of discretization from finite Euler steps or a consequence of NLL-likes over high dimensional joints, its worth…