CorrAdapter-Model-on-MVAdapter

This model repository contains the trained MVAdapter + CorrAdapter* checkpoint for image-conditioned static multi-view generation from the CorrAdapter project.

CorrAdapter* is the trained variant of CorrAdapter on top of MV-Adapter image-to-multiview SDXL. The checkpoint is saved in Hugging Face sharded PyTorch format:

pytorch_model.bin.index.json
pytorch_model-00001-of-00004.bin
pytorch_model-00002-of-00004.bin
pytorch_model-00003-of-00004.bin
pytorch_model-00004-of-00004.bin

This is not a standalone Diffusers pipeline. Use it with the CorrAdapter MVAdapter subproject.

Usage

From the CorrAdapter MVAdapter directory:

python -m scripts.inference_i2mv_sdxl \
  --adapter_path SuhZhang/CorrAdapter-Model-on-MVAdapter \
  --adapter_weight_name pytorch_model.bin.index.json \
  --image assets/demo/i2mv/A_decorative_figurine_of_a_young_anime-style_girl.png \
  --text "A decorative figurine of a young anime-style girl" \
  --seed 21 \
  --output outputs/demo_i2mv_trained/output.png \
  --remove_bg

For local loading, replace --adapter_path with the local model directory path.

Expected Code

The loader in mvadapter/loaders/custom_adapter.py supports pytorch_model.bin.index.json and loads only the keys needed by the MVAdapter CorrAdapter pipeline. The inference script also accepts:

--adapter_weight_name auto

when the directory contains pytorch_model.bin.index.json.

Citation

@InProceedings{Zhang_2026_CVPR,
  author = {Zhang, Shihua and Shen, Qiuhong and Wang, Xinchao},
  title = {Align Images Before You Generate},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2026},
  pages = {30521-30531}
}

Please also cite MV-Adapter when using this checkpoint.