DWD Lets Diffusion Transformers Predict Pixels Without a Decoder
A research preprint introduces Decoding Without a Decoder (DWD), which adds a linear head to a fine-tuned diffusion transformer so it predicts final pixels directly while the original head continues to guide denoising. The authors report maintained or improved pixel quality across latent image and video generators, a 23% reduction in warping error for Wan video generation, and an ImageNet FID improvement from 1.55 to 1.43 in pixel-flow text-to-image generation.
Removing the separate decoder could reduce the complexity and inference cost of latent generative pipelines. The authors also report improved pixel quality, along with reduced detail distortion and flickering in Wan video generation.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.22287 (cs)
[Submitted on 17 Nov 2026 (v1), last revised 18 Nov 2026 (this version, v2)]
Title:Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers
Authors:Hansheng Chen, Shengqu Cai, Jan Ackermann, Rayhan Zirvi, Gordon Wetzstein, Xinchen Yan, Leonidas Guibas
View a PDF of the paper titled Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers, by Hansheng Chen and 6 other authors