adrianrm/ambient-o-noise-classifier-blur06-prob05-iter15k

Ambient Diffusion Omni (Ambient-o): Training Good Models with Bad Data

Dataset Description

Ambient Diffusion Omni (Ambient-o) is a framework for using low-quality, synthetic, and out-of-distribution images to improve the quality of diffusion models. Unlike traditional approaches that rely on highly curated datasets, Ambient-o extracts valuable signal from all available images during training, including data typically discarded as "low-quality."

This model card is for the noise classifier on cifar10 for blurring corruptions with sigma_B=0.6, used in our pixel diffusion experiments based on EDM. The architecture is a time-dependent half U-Net backbone + linear projection.

Given a noisy image x_t and a noise level sigma_t, it predicts the probabily that the un-noised image x is blurry or not blurry. We use this to "annotate" the blurry (low-quality) part of our dataset with the minimum noise level sigma_tn at which its safe to use.

Visualization

For visualization purposes, we include here a figure from our paper explaining how it works:

As shown in the figure above, when fed the blurry image our annotation pipeline would output 1.5.

Usage instructions

You can load this using the following code, after cloning our repository and running cd pixel-diffusion

import dnnlib
from huggingface_hub import hf_hub_download
import json
from safetensors.torch import load_file

# Get network args from HF
options_path = hf_hub_download(repo_id=ckpt_file, filename="training_options.json")
options = json.load(open(options_path, "r", encoding="utf-8"))

interface_kwargs_path = hf_hub_download(repo_id=ckpt_file, filename="interface_kwargs.json")
interface_kwargs = json.load(open(interface_kwargs_path, "r", encoding="utf-8"))

# Construct architecture
net = dnnlib.util.construct_class_by_name(**options['network_kwargs'], **interface_kwargs)

# Load state dict
state_dict_path = hf_hub_download(repo_id=ckpt_file, filename="ema.safetensors")
state_dict = load_file(state_dict_path)
net.load_state_dict(state_dict)
net.eval()

Citation

@article{daras2025ambient,
  title={Ambient Diffusion Omni: Training Good Models with Bad Data},
  author={Daras, Giannis and Rodriguez-Munoz, Adrian and Klivans, Adam and Torralba, Antonio and Daskalakis, Constantinos},
  journal={arXiv preprint},
  year={2025},
}