we do not have a full checkpoint conversion validation, if you encounter pipeline loading failure and unsidered output, please contact me via bili_sakura@zju.edu.cn

BiliSakura/UNSB-ckpt

Converted UNSB (Unpaired Neural Schroedinger Bridge) generator checkpoints packaged for easy use with pytorch-image-translation-models.

Model variants

Model	Direction
horse2zebra	horse -> zebra
label2cityscape	semantic label -> cityscape
male2female	male -> female
mat2sat	map -> satellite
summer2winter	summer -> winter

Repository layout

{model_name}/
  generator/
    config.json
    diffusion_pytorch_model.safetensors

Usage

from PIL import Image

from src.pipelines.unsb import UNSBPipeline

pipe = UNSBPipeline.from_pretrained(
    "/path/to/UNSB-ckpt/horse2zebra",
    subfolder="generator",
    device="cuda",
    scheduler_num_timesteps=5,
    scheduler_tau=0.01,
)

source = Image.open("/path/to/source.png").convert("RGB")
out = pipe(source_image=source, output_type="pil")
out.images[0].save("unsb_output.png")

Conversion note

• Source checkpoints were loaded from latest_net_G.pth.
• Parameter keys were remapped to match src.models.unsb.UNSBGenerator naming in pytorch-image-translation-models:
  • Dense_time -> dense_time
  • legacy residual-block style.* parameters removed
• Converted weights were validated with strict state-dict loading.

Citation

@inproceedings{kim2024unpaired,
  title={Unpaired Image-to-Image Translation via Neural Schroedinger Bridge},
  author={Kim, Dongjun and others},
  booktitle={ICLR},
  year={2024}
}