DWD paper removes the decoder from diffusion models, cuts cost and hits new SOTA
Crazy-Repeat-2006 · reddit · 2026-11-19
A new paper introduces Decoding Without a Decoder (DWD): a finetuned diffusion transformer predicts final pixels directly through a linear head, eliminating the decoder in latent models and cutting inference cost.
Method:
- Role separation: the original denoising head keeps driving sampling (anchored to the pretrained model to preserve trajectories), while a new DWD head predicts pixels only at the final step
- Efficient extension: only the last few transformer blocks are finetuned, trained with pixel-quality losses
Results:
- Across latent image/video generators, DWD removes the decoder and lowers inference cost while maintaining or improving quality; on Wan video it cuts warping error by 23%, visibly reducing detail distortion and flickering
- The same trick improves pixel diffusion models too: better texture detail in text-to-image, and ImageNet FID improved from 1.55 to 1.43 — a new SOTA among pixel flow models
Related event: DWD Paper Drops Decoder, Diffusion Model Predicts Pixels Directly(2 posts)→
More from Multimodal
- HiT image generator learns its own abstraction hierarchy, hits 1.48 FID on ImageNet 256 — LucaAmb · 2026-11-19
- Osmo rebrands as applied AI lab, launches Posty video agent that cuts footage in under 30 min — yawnxyz · 2026-11-19
- MOSS-Image-Base nails small real-world sign text: lost cat flyers and passive-aggressive notes — Impressive-Kick293 · 2026-11-19
- Dev ports MiniMax H3 to Swift, cuts 480p video gen to ~4 min on a 16GB M1 MacBook — Lokoto123 · 2026-11-19
- Dreamina Black Friday: unlimited Seedance 2.5 video for 30 days, credits up to 83% off — FellMentKE · 2026-11-19
- How I built a music demo with Gemini 4 and Antigravity — alexanderchen · 2026-11-19