NVIDIA researchers introduce PixelDiT pixel diffusion model

openaccess.thecvf.com19 min read

Researchers propose PixelDiT, a single-stage, end-to-end diffusion transformer model that denoises images directly in pixel space without relying on a separate autoencoder. The model uses a dual-level architecture featuring patch-level DiT blocks for global semantics and pixel-level PiT blocks with pixel-wise AdaLN modulation and token compaction for texture details. Experimental results show that PixelDiT achieves competitive generative performance on ImageNet and text-to-image benchmarks while avoiding VAE reconstruction artifacts. Open article →