DiT-XL/2 (256x256)

https://arxiv.org/abs/2212.09748

Lucid port of https://huggingface.co/facebook/DiT-XL-2-256, converted to Lucid-native safetensors.

Available weights

Usage

import lucid.models as models
from lucid.models.weights import DiTXLarge2Weights

# default tag
model = models.dit_xlarge_2(pretrained=True)

# explicit tag (enum or string)
model = models.dit_xlarge_2(weights=DiTXLarge2Weights.IMAGENET1K_256)
model = models.dit_xlarge_2(pretrained="IMAGENET1K_256")

# preprocessing travels with the weights
weights = DiTXLarge2Weights.IMAGENET1K_256
preprocess = weights.transforms()
out = model(preprocess(image)[None])
# A latent diffusion backbone predicts noise, not labels.
eps = out[:, : model.config.in_channels]  # (B, C, H, W)

Conversion

Converted from https://huggingface.co/facebook/DiT-XL-2-256 via python -m tools.convert_weights dit_xlarge_2 --tag IMAGENET1K_256. Key mapping + numerical parity verified against the source.

License

cc-by-nc-4.0 — inherited from the original weights.

Citation

@inproceedings{peebles2023scalable,
  title={Scalable Diffusion Models with Transformers},
  author={Peebles, William and Xie, Saining},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages={4195--4205},
  year={2023}
}