MMLab CUHK / Shanghai AI Lab · Text-to-Image

NiT (Native-resolution diffusion Transformer)

A diffusion transformer architecture that generates images at arbitrary resolutions and aspect ratios using dynamic tokenization, variable-length sequence processing, and 2D structural priors. Achieves SOTA performance on ImageNet benchmarks (2.03 FID at 256×256, 1.45 FID at 512×512) with strong zero-shot generalization to unseen resolutions up to 2048×2048.