PixRestore: Unified Image Restoration via Pixel Diffusion Transformer
PixRestore restores degraded images in pixel space instead of leaning on latent text-to-image diffusion.
The paper says latent diffusion VAEs can lose details that matter for restoration and can add content-inconsistent artifacts. PixRestore uses a VAE-free Diffusion Transformer trained from scratch, with flow matching applied directly to patchified pixels. It conditions the model by estimating which feature layers are reliable under each degradation, then fuses or supervises them accordingly. The authors report best overall fidelity, perceptual quality, and robustness against competing unified restoration models, with about 50M parameters and single-step inference. HF Daily Papers' note
The paper says latent diffusion VAEs can lose details that matter for restoration and can add content-inconsistent artifacts. PixRestore uses a VAE-free Diffusion Transformer trained from scratch, with flow matching applied directly to patchified pixels. It conditions the model by estimating which feature layers are reliable under each degradation, then fuses or supervises them accordingly. The authors report best overall fidelity, perceptual quality, and robustness against competing unified restoration models, with about 50M parameters and single-step inference. HF Daily Papers' note
score 4