PixArt-Sigma

Generating images directly at up to 4K resolution, PixArt-Sigma is a text-to-image Diffusion Transformer developed by researchers from Huawei Noah's Ark Lab together with DLUT, HKU, and HKUST. Despite that output range the transformer holds only about 0.6 billion parameters, an efficiency the team attributes to weak-to-strong training, a strategy that evolves the model from its PixArt-alpha predecessor using higher quality data rather than training from scratch. Compared with PixArt-alpha it extends the T5 text token length from 120 to 300 for richer prompts and swaps in the SDXL VAE, supporting resolutions from 256 up through 2K and 4K across varied aspect ratios. Training and inference code, checkpoints at multiple resolutions, a Gradio demo, and Hugging Face Diffusers integration are all released openly. The model draws researchers studying efficient diffusion transformers and practitioners who want high-resolution generation on modest compute budgets.