Most diffusion baselines conflate denoising with understanding, says kalomaze

kalomaze · x · 2026-07-23

99% of public diffusion baselines, he argues, incorrectly tie together two separate roles: the network that denoises and the network that does “deep/nontrivial understanding.”

His suggestion is to condition a cheap denoiser on deterministic vectors from a larger model, leaving the expensive reasoning upstream. He says this leaves “aggressive headroom,” even before adding MoE. He also adds a pointed critique of diffusion research: the field gets away with this because the dominant benchmark culture is still a meme, centered on proxy metrics like ImageNet FID that can be gamed.

Related event: Rethinking Diffusion: Denoising and Understanding Should Decouple(4 posts)→

Original post →

More from Research

Research channel →