Stanford DWD paper: linear head predicts pixels directly, ditching the diffusion decoder
kwangmoo_yi · x · 2026-11-19
Researchers from Stanford, Luma AI, and Caltech introduce DWD (Decoding Without a Decoder), letting a finetuned diffusion transformer produce final pixels via a linear head, removing the decoder from latent diffusion pipelines.
Key ideas:
- Separates roles: the original denoising head keeps driving sampling, while a new DWD head predicts pixels at the final step.
- Extending pretrained latent models only requires finetuning the last few transformer blocks, training the DWD head with pixel quality losses while anchoring the denoising head to preserve the sample trajectory.
Results:
- Across latent image and video generators, removing the decoder maintains or improves pixel quality at lower inference cost; on Wan video it reduces detail distortion and flickering, cutting warping error by 23%.
- It also improves pixel diffusion models without decoders: better texture detail in text-to-image and ImageNet FID improved from 1.55 to 1.43, a new SOTA among pixel diffusion models.
Code is coming soon; the authors suggest the V(Q)AE used in training may have been just a speedup that can be discarded entirely.
Related event: DWD Paper: Diffusion Models Predict Pixels Without a Decoder(4 posts)→
More from Multimodal
- AI video brings Mona Lisa, Van Gogh and Dalí together across centuries — sebpaquet · 2026-11-19
- Fully Seedance 2.5-generated short shows AI video getting comedically coherent — anthara_ai · 2026-11-19
- Kling Image 4.0 debuts at #7 on Artificial Analysis, only model with native 8K output — anthara_ai · 2026-11-19
- Creator explains his slopcore video workflow built on a Claude plugin — AIandDesign · 2026-11-19
- Blender Bench: frontier models drive real Blender via MCP, humans blind-vote on 3D scenes — Gruku · 2026-11-19
- Core Memory profiles Ari Kuschnir, the creator behind viral AI parody videos — kyliebytes · 2026-11-19