PIXART-α

PIXART-α is a Transformer-based text-to-image diffusion model designed for photorealistic image synthesis with significantly reduced training costs. It achieves competitive image quality with state-of-the-art generators like Imagen, SDXL, and Midjourney, while requiring only 10.8% of Stable Diffusion v1.5's training time (~675 vs. ~6,250 A100 GPU days), slashing costs to approximately $26,000 and reducing CO2 emissions by 90%. The model incorporates three core innovations: training strategy decomposition (separately optimizing pixel dependency, text-image alignment, and aesthetic quality), an efficient T2I Transformer with cross-attention modules, and high-informative data labeling using a large vision-language model for dense pseudo-captions. PIXART-α supports high-resolution synthesis up to 1024px and can be extended with ControlNet and DreamBooth for customized generation.