Pixel-Space Diffusion Models

Pixel-space diffusion models learn to generate or reconstruct images directly in RGB, aiming to preserve details that image tokenizers or autoencoders can discard. Common approaches include Diffusion Transformers and flow matching, with multiscale or frequency-aware denoising, semantic representation guidance, and token compression or distillation to reduce computation. On benchmarks such as ImageNet at 256 or 512 pixels and text-to-image evaluations such as GenEval, pixel-space models now reach competitive quality, and the same techniques are being applied to restoration, geometry, and inverse problems. The main tension is that raw pixels demand more computation and can make training slower than latent-space methods, while eliminating the latent bottleneck does not always improve downstream results.