DWD paper: drop the decoder — linear pixel head sets pixel-SOTA FID 1.43 on ImageNet

GordonWetzstein · x · 2026-11-19

New work from Gordon Wetzstein's group asks whether a diffusion transformer needs a separate decoder. DWD adds a linear head on DiT that reads pixels directly from final features while preserving the denoising trajectory. Results: same or better quality across every tested latent model, less flicker in Wan video; surprisingly, the separation also improves direct pixel generators, cutting ImageNet FID from 1.55 to 1.43 (pixel SOTA) with photorealistic textures in T2I.

Related event: DWD Paper Drops Decoder, Diffusion Model Predicts Pixels Directly(2 posts)→

Original post →

More from Multimodal

Multimodal channel →