ComboStoc-XL-2-256x256

ComboStoc-XL/2 (256×256, UNSYNC_ALL @ 800K)

Re-hosted pre-trained checkpoint for ComboStoc: Combinatorial Stochasticity for Diffusion Generative Models (ACM Transactions on Graphics, SIGGRAPH 2026).

• Paper: https://arxiv.org/abs/2405.13729
• Project page: https://ruixu.me/html/ComboStoc/index.html
• Original code: https://github.com/(ComboStoc authors) — this repo mirrors the weights originally distributed via the authors' Google Drive.
• Authors: Rui Xu, Jiepeng Wang, Hao Pan, Yang Liu, Xin Tong, Shiqing Xin, Changhe Tu, Taku Komura, Wenping Wang.

What this is

This is the ComboStoc-XL/2 image diffusion model (SiT-style transformer, XL depth, patch size 2) trained on class-conditional ImageNet 256×256 with the UNSYNC_ALL combinatorial-stochasticity setting, at 800K training steps. The checkpoint file is ALL_0800000.pt.

ComboStoc studies the combinatorial complexity of diffusion training: the space spanned by combinations of dimensions and attributes is under-sampled by standard training. ComboStoc constructs stochastic processes that fully exploit these combinatorial structures, accelerating training and enabling asynchronous per-dimension/per-attribute time steps at test time.

Files

• ALL_0800000.pt — model checkpoint. Follows the SiT checkpoint convention; the dict contains an ema entry which is the recommended weights for sampling.

Usage

Load with huggingface_hub.hf_hub_download and the ComboStoc model definition (ComboStoc-XL/2, num_classes=1000, image_size=256) plus the stabilityai/sd-vae-ft-mse VAE. See the accompanying demo Space for a full example.

License

GNU AGPL-3.0, matching the original ComboStoc source repository. If you use these weights, you must comply with AGPL-3.0 terms.

Citation

@misc{xu2024combostoccombinatorialstochasticitydiffusion,
      title={ComboStoc: Combinatorial Stochasticity for Diffusion Generative Models},
      author={Rui Xu and Jiepeng Wang and Hao Pan and Yang Liu and Xin Tong and Shiqing Xin and Changhe Tu and Taku Komura and Wenping Wang},
      year={2024},
      eprint={2405.13729},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2405.13729},
}

These weights are a community re-host to enable a Hugging Face Space demo. All credit for the model belongs to the original authors.