PixArt
PixArt-alphaPixArt is a family of open text-to-image diffusion transformers from the PixArt-alpha research team. PixArt-α (2023) was followed by PixArt-Σ (Sigma, 2024), which uses a T5 text encoder and builds on its predecessor's weights to generate images at up to 4K resolution.
Dormant: no release since the PixArt-Σ 2K checkpoint of April 2024, and both repositories were last updated in October 2024. Diffusers supports the models natively.
Openness
3 high confidence- weights
- open (PixArt-Σ checkpoints download from the Hub without a gate)
- data
- described (only a toy dataset is released for the training code; the full training set is not)
- code
- open (training and inference code, including LoRA and one-step DMD training, is published)
- license
- CreativeML-OpenRAIL++-M (the PixArt-Σ weights, per the model card)+Apache-2.0 (the code)
PixArt-Σ's weights are under the CreativeML Open RAIL++-M license, which allows commercial use with restrictions on harmful uses, and the code is Apache 2.0. Training code is published with a toy dataset; the full training set is not released.
- https://huggingface.co/api/models/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS?expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=createdAt&expand[]=siblings recorded 2026-11-12
Hugging Face API record for PixArt-alpha/PixArt-Sigma-XL-2-1024-MS: license openrail++, gated false, transformer, VAE and T5 safetensors.
- https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/raw/main/README.md recorded 2026-11-12
Card front matter "license: openrail++" and "License: CreativeML Open RAIL++-M License"; a Transformer latent diffusion model with "one fixed, pretrained text encoders (T5)".
- https://ungh.cc/repos/PixArt-alpha/PixArt-sigma/files/master/LICENSE recorded 2026-11-12
LICENSE body: Apache License, Version 2.0.
- https://ungh.cc/repos/PixArt-alpha/PixArt-sigma/files/master/README.md recorded 2026-11-12
README: "Mar. 29, 2024. PixArt-Σ training & inference code & toy data are released"; "Apr. 24, 2024. LoRA code is released"; the training section starts from "the toy dataset".
Adoption
2 high confidenceAdoption is measured as Hugging Face downloads summed across the PixArt-alpha checkpoints, three quarters of them PixArt-Σ at 1024px.
- https://huggingface.co/api/models?author=PixArt-alpha&sort=downloads&direction=-1&limit=100 recorded 2026-11-12
Listing of the 14 PixArt-alpha repositories sorted by downloads: PixArt-Sigma-XL-2-1024-MS 12469, PixArt-XL-2-512x512 1382, PixArt-XL-2-1024-MS 1263, pixart_sigma_sdxlvae_T5_diffusers 1107, the rest under 100 each; 16,327 in total.
Capability
1 medium confidencePixArt has no placement on the public text-to-image arena, which now lists only current models, while FLUX.2 [dev] sits in its upper half.
- https://artificialanalysis.ai/image/leaderboard/text-to-image recorded 2026-11-12
The overall text-to-image board as served lists 112 models, from GPT Image 2.6 Sunburst first to Bagel 112th, and no PixArt entry.
Verified 2026-11-12