Stanford removes the decoder: direct pixel prediction from diffusion transformers cuts cost, lifts quality

StanfordUniversity · hf · 2026-11-19

Stanford (with Luma AI and Caltech) presents "Decoding Without a Decoder": by adding a final-step prediction head to a finetuned diffusion transformer, the method predicts pixels directly from the DiT, removing the decoder from latent diffusion models.

The approach reduces inference cost while improving pixel-level quality in both image and video generation; paper figures show the decoder-free DWD results beating baselines on Wan2.1 14B video latent diffusion and T2I pixel diffusion (e.g., sharper signage text).

Original post →

More from Multimodal

Multimodal channel →