InvSR Model Card

This model card focuses on the models associated with the InvSR project, which is available here.

Model Details

• Developed by: Zongsheng Yue

• Model type: Arbitrary-steps Image Super-resolution via Diffusion Inversion

• Model Description: This is the model used in Paper.

• Resources for more information: GitHub Repository.

• Cite as:

  @article{yue2024invSR,
    author    = {Zongsheng Yue, Kang Liao, Chen Change Loy},
    title     = {Arbitrary-steps Image Super-resolution via Diffusion Inversion},
    journal   = {arXiv preprint arXiv:2412.09013},
    year      = {2024},
  }
  

Limitations and Bias

Limitations

• InvSR requires a tiled operation for generating a high-resolution image, which would largely increase the inference time.
• InvSR sometimes cannot keep 100% fidelity due to its generative nature.
• InvSR sometimes cannot generate perfect details under complex real-world scenarios.

Bias

While our model is based on a pre-trained SD-Turbo model, currently we do not observe obvious bias in generated results.

Training

Training Data The model developer used the following dataset for training the model:

• Our model is finetuned on LSDIR + 20K samples from FFHQ datasets.

Training Procedure InvSR achieves the goal of image super-resolution via diffusion inversion technique on SD-Turbo, detailed training pipelines can be found in our GitHub repo.

We currently provide the following checkpoints:

• noise_predictor_sd_turbo_v5.pth: Noise estimation network trained for SD-Turbo.

Evaluation Results

See Paper for details.