Research map / Image generation / Fast diffusion sampling
Ambient Diffusion: Learning Clean Distributions from Corrupted Data
Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alex Dimakis, Adam Klivans
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add further corruption (extra masking) to training inputs and compute loss on all originally observed pixels (ablation) | Ambient Diffusion · CIFAR-10 high corruption | FID | Improved |
| train on corrupted data with Ambient Diffusion vs AmbientGAN | CIFAR-10 generation · CIFAR-10 all corruption levels | generation quality (FID/Inception) | Improved |
| single-step Ambient Diffusion restoration vs DPS (100 NFE) at p=0.6 | CelebA-HQ inpainting · CelebA-HQ random inpainting p=0.6 | PSNR: 28.21 -> 31.51 | Improved |
| single-step Ambient Diffusion restoration vs DPS (100 NFE) at p=0.4 | AFHQ inpainting · AFHQ random inpainting p=0.4 | PSNR: 34.06 -> 33.27 | Hurt |
| Ambient Diffusion vs supervised MAE inpainting | AFHQ inpainting · AFHQ random inpainting p=0.4 | PSNR: 28.88 -> 33.27 | Improved |
| add Reconstruction Guidance term to Fixed Mask sampler (ablation) | Ambient Diffusion sampler · CIFAR-10 FID | FID: 11.70 -> 11.59 (p=0.2); 46.27 -> 46.31 (p=0.8) | Mixed |
| fine-tune Deepfloyd IF on 80% pixel-deleted CelebA images vs standard fine-tuning on clean images (ablation) | Deepfloyd IF fine-tuning · CelebA 3000-image subset | memorization (DINOv2 nearest-neighbor similarity) | Improved |
| increase corruption level in training data (ablation) | Ambient Diffusion · CelebA full dataset | memorization vs sample quality | Mixed |
Closest papers by meaning
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.92)
- Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025 · similarity 0.91)
- Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS 2025 · similarity 0.91)
- An Expectation-Maximization Algorithm for Training Clean Diffusion Models from Corrupted Observations (NeurIPS 2024 · similarity 0.89)
- Stochastic Forward–Backward Deconvolution: Training Diffusion Models with Finite Noisy Datasets (ICML 2025 · similarity 0.89)
- How Much is a Noisy Image Worth? Data Scaling Laws for Ambient Diffusion. (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
