Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers

arxiv.org1 min read

Researchers introduce Decoding Without a Decoder (DWD), a method that enables finetuned diffusion transformers to produce final pixels directly through a linear head rather than relying on a separate decoder. By separating the denoising sampling trajectory from the final pixel prediction step, DWD removes decoder complexity and reduces inference cost while maintaining or improving pixel quality across image and video generators. Open article →