PixArt-Sigma
A local text-to-image model with 2K and 4K generation, a Gradio browser interface, and support for PyTorch and Hugging Face Diffusers.
PixArt-Sigma generates images from text prompts on your own machine or a self-hosted server. It's for developers and researchers who want downloadable image models and the code to train their own.
- High-resolution images: Generate at 2K and 4K resolutions, with support for longer prompts than PixArt-α.
- Browser interface: Use a Gradio front end for image generation while the model runs locally or on your server.
- PyTorch and Diffusers: Run image generation through PyTorch or Hugging Face Diffusers, with support for converting PyTorch .pth checkpoints to Diffusers format.
- Model training: Train models with the supplied code without a separate feature-extraction stage, or use pretrained checkpoints that include 1024-pixel and 2K models.
PixArt-Sigma builds on PixArt-α's diffusion transformer architecture and uses a T5 text encoder with the SDXL variational autoencoder. Its Python code is open source under Apache 2.0.
Logotyper
