Reproduce “Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hansheng Chen, Shengqu Cai, Jan Ackermann, Rayhan Zirvi, Gordon Wetzstein, Xinchen Yan, Leonidas Guibas

DWD adds a separate linear pixel head to diffusion transformers, leaving the denoising head to drive sampling. In latent generators, it predicts pixels at the final step and bypasses the decoder; in pixel diffusion, it separates pixel-quality objectives from denoising. On ImageNet, FID falls from 1.55 to 1.43, while a shared-head ablation reaches 1.71. These gains are reported in tested models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.