Skip to content

khaledkah/tv-snr-diffusion

Repository files navigation

Public source code for our paper: TV/SNR - Enhancing Diffusion Models Efficiency by Disentangling Total-Variance and Signal-to-Noise Ratio

While the repository is currently being refactored, we will provide the notebook to generate our toy examples under analytic_score_toy_example.ipynb. We added the code for our method to the tv_snr folder. The refactored version will be made public in the coming days.

The computer vision code is based on the original EDM repo and can be used in the same way using generate_tv_snr.py to generate samples.

Example command to generate images using our best method:

python generate_tv_snr.py --outdir=out --snr_schedule sig3 --linear_time --num_steps 64 --tau 1.0 \
    --network=https://nvlabs-fi-cdn.nvidia.com/edm/pretrained/baseline/baseline-cifar10-32x32-uncond-vp.pkl --grid

How to cite

if you use TV/SNR in your research, please cite the corresponding publication:

Kahouli, K., Ripken, W., Gugler, S., Unke, O. T., Müller, K. R., & Nakajima, S. (2025). Enhancing Diffusion Models Efficiency by Disentangling Total-Variance and Signal-to-Noise Ratio. arXiv preprint arXiv:2502.08598.

@article{kahouli2025enhancing,
  title={Enhancing Diffusion Models Efficiency by Disentangling Total-Variance and Signal-to-Noise Ratio},
  author={Kahouli, Khaled and Ripken, Winfried and Gugler, Stefan and Unke, Oliver T and M{\~A}{\v{z}}ller, Klaus-Robert and Nakajima, Shinichi},
  journal={arXiv preprint arXiv:2502.08598},
  year={2025}
}

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published