Model Card for ambient_laws_celeba_sigma_0.05_corruption_0.7_keep_1.0

General Information 📚

This model is part of a collection of models that were trained for the paper: How Much is a Noisy Image Worth? 👀.

In this paper, we show that noisy images can be very useful in training diffusion generative models, as long as a small set of clean images is available.

How to use this model 🚀

Detailed instructions are in our GitHub repository.

You can clone the repository with the following command:

git clone https://github.com/giannisdaras/ambient-laws.git

and you can use the following function to load the model from the hub:

import dnnlib
import json
from huggingface_hub import hf_hub_download

def load_hf_checkpoint(repo_id):
    config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
    model_config = json.load(open(config_path, "r", encoding="utf-8"))
    model_config['class_name'] = 'training.networks.EDMPrecond'
    net = dnnlib.util.construct_class_by_name(**model_config)
    net = net.from_pretrained(repo_id)
    return net 

model = load_hf_checkpoint("giannisdaras/ambient_laws_celeba_sigma_0.05_corruption_0.7_keep_1.0")

Model Description 📝

This model was trained on celeba using 100.0% of the samples in the dataset.

From the samples kept, 70.0% of them were clean images and 30.0% of them were noisy images at noise level sigma=0.05.

The model was trained for a total of 100000 training steps.

Citation 📄

If you find this work useful, please consider citing the following paper:

@article{daras2024imageworth,
  author    = {Giannis Daras and Yeshwanth Cherapanamjeri and Constantinos Daskalakis},
  title     = {How much is a noisy image worth? Data scaling laws for Ambient Diffusion.},
  journal   = {arXiv preprint arXiv:2411.02780},
  year      = {2024},
  url       = {https://arxiv.org/abs/2411.02780}
}

This model was shared by @giannisdaras.