Skip to content

PixArt

Model Source Revision Weight license Parity (CPU, fp32)
pixart-sigma-xl-1024 PixArt-alpha/PixArt-Sigma-XL-2-1024-MS e102b35 CreativeML-OpenRAIL++-M Verified
pixart-sigma-xl-512 PixArt-alpha/PixArt-Sigma-XL-2-512-MS 76fb7eb CreativeML-OpenRAIL++-M Verified
pixart-sigma-xl-256 PixArt-alpha/PixArt-Sigma-XL-2-256x256 495cb1b OpenRAIL (exact terms unspecified) Verified
pixart-sigma-xl-2k PixArt-alpha/PixArt-Sigma-XL-2-2K-MS cd274f4 OpenRAIL (exact terms unspecified) Verified
pixart-alpha-xl-1024 PixArt-alpha/PixArt-XL-2-1024-MS b89adad CreativeML-OpenRAIL++-M Verified
pixart-alpha-xl-512 PixArt-alpha/PixArt-XL-2-512x512 50f7021 CreativeML-OpenRAIL++-M Verified
pixart-lcm-xl-1024 PixArt-alpha/PixArt-LCM-XL-2-1024-MS e153060 CreativeML-OpenRAIL++-M Verified

Catalog: pixart-sigma, pixart-alpha · Models: pixart-sigma-xl-1024, pixart-sigma-xl-512, pixart-sigma-xl-256, pixart-sigma-xl-2k, pixart-alpha-xl-1024, pixart-alpha-xl-512, pixart-lcm-xl-1024 · Task: t2i

Architecture. A T5-v1.1-XXL text encoder supplies up to 300 tokens for Sigma and 120 for Alpha to a 0.6B diffusion transformer with 28 blocks, width 1152 and 16 attention heads of width 72. The transformer uses shared ada_norm_single timestep modulation, caption projection, a 2D sine-cosine position embedding and learned sigma prediction. Sigma uses interpolation scale 2 and the SDXL VAE with scaling factor 0.13025. Alpha adds resolution and aspect-ratio embeddings and uses the SD VAE with scaling factor 0.18215. Both use the order-2 midpoint DPM-Solver++ multistep scheduler and bin requested resolutions to their supported aspect-ratio tables. PixArt-LCM reuses the Alpha layout with an LCM scheduler and its own distilled transformer.

Usage.

python -m imagehub.pretrained pixart-sigma-xl-1024 checkpoints/pixart-sigma-xl-1024
from imagehub import pipeline

with pipeline(
    "t2i", backend="pixart-sigma-xl-1024", model="checkpoints/pixart-sigma-xl-1024", device="mps"
) as pipe:
    result = pipe("a cat, oil painting", seed=42, negative_prompt="blurry")
    result.images[0].save("cat.png")

Resolutions. Each resolution is a separate checkpoint with its own transformer: Sigma at 256, 512, 1024 and 2K (interpolation scales 0.5, 1, 2 and 4) and Alpha at 512 and 1024. The Sigma 256, 512 and 2K repositories publish only a transformer, so python -m imagehub.pretrained fetches the text encoder, tokenizer, VAE and scheduler from the Sigma 1024 repository, as the model cards describe. The Alpha resolution and aspect-ratio embeddings exist only in the 1024 model. Resolution binning uses the table that matches the transformer (256, 512, 1024 or 2048).

Settings. PixArt-LCM defaults to 4 steps, guidance 0.0 and 1024×1024 output; all other models default to 20 steps and guidance 4.5, and to a square image of their native size (256, 512, 1024 or 2048 pixels). Sigma defaults to a maximum sequence length of 300 and Alpha and LCM to 120. The other supported options are seed, latents, negative_prompt, use_resolution_binning and max_sequence_length. Unrecognized settings raise an error.

Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.

Out of scope. PixArt-DMD, ControlNet, fp16/bf16 parity and caption cleaning are not supported. Caption cleaning requires bs4 and ftfy; parity is against clean_caption=False, which diffusers uses when those packages are unavailable.

License. Both models use the CreativeML-OpenRAIL++-M license. imagehub does not distribute weights.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.