PixArt-α is a free, open-source text-to-image model from Huawei Noah’s Ark Lab and collaborators that brought Diffusion Transformer (DiT) image generation to the community at a fraction of the usual training cost. It produces photorealistic images up to 1024×1024 resolution while training dramatically faster and cheaper than comparable models — a result that influenced much of the DiT research that followed.
PixArt-α is an open-source project on GitHub with about 3,300 stars, released in 2023–2024. Its key contribution is an efficient training recipe: a carefully curated dataset pipeline, a transformer decomposed into cross-attention layers for text conditioning, and a three-stage training strategy that brought total training cost down to a small fraction of models like Stable Diffusion v1.5, while matching or exceeding their image quality and text alignment. The repository includes model weights, training and inference code, and the PixArt-Σ successor line continued the work. Despite its research origins, PixArt-α remains a fast, capable open image generator that runs on a single consumer GPU.
PixArt-α is completely free and open source under the Apache-2.0 license. Download the weights and generate images on your own hardware at no cost.
The model conditions on T5 text embeddings and supports classifier-free guidance for prompt control, and the released codebase documents the full data-curation and training pipeline, making PixArt-α as valuable as a study in efficient model training as it is as an image generator in its own right.
PixArt-α is for researchers and developers interested in efficient diffusion-transformer training, and for creators who want a fast, high-quality open text-to-image model to run or fine-tune locally.