Research map / Image generation / Fast diffusion sampling
Ambient Dataloops: Generative Models for Dataset Refinement
Adrian Rodriguez-Munoz, William Daspit, Adam Klivans, Antonio Torralba, Constantinos Costis Daskalakis, Giannis Daras
Read the paperCheck your idea against 137,217 papers
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 one Ambient Dataloops restoration loop on top of Ambient Omni | Ambient Omni (loop 0) · CIFAR-10 with 90% corrupted data (blur and JPEG) | unconditional FID: 5.689 -> 4.947 (sigma_B=0.6); up to 17% FID reduction | Improved |
| treat corrupted data with ambient noise-level training instead of filtering or using all data as clean | CIFAR-10 90% corrupted · CIFAR-10 blur sigma_B=0.6 | FID: 11.26 (no filtering), 8.79 (filtering) -> 5.689 (Ambient Omni) | Improved |
| add a second restoration pass using the loop-1 model (more loops) (ablation) | Ambient Dataloops · CIFAR-10 blur sigma_B=0.6 | conditional metrics | Mixed |
| run 3 loops, reducing dataset noise by factor 8 each time (ablation) | Ambient Dataloops · CIFAR-10 blur sigma_B=0.6 | FID | Improved |
| denoise dataset too fast or too slow, or beyond the model's capacity (ablation) | Ambient Dataloops · CIFAR-10 blur | FID | Hurt |
| partially restore DiffusionDB to sigma 1.0 (half of 2.0) and retrain | Ambient Omni text-to-image model · COCO zero-shot | FID-30K: 10.61 -> 10.06 | Improved |
| partially restore DiffusionDB and retrain | Ambient Omni text-to-image model · COCO zero-shot | Clip-FD-30K: 9.40 -> 8.83 | Improved |
| partially restore DiffusionDB and retrain | Ambient Omni text-to-image model · GenEval overall | score: 0.47 -> 0.47 | No clear effect |
| increase SIMS guidance strength to improve image quality | SIMS guidance baseline · COCO zero-shot | FID: 13.03 -> 13.23 (omega 0.0 -> 0.5) | Hurt |
| restore bottom 20% of ImageNet with generative model | ImageNet dataset · ImageNet quality metrics | Aesthetic score: 4.03 -> 4.41 | Improved |
| finetune ImageNet generator on restored ImageNet | Degeorge et al. 2025 model · ImageNet generation | ImageReward, HPSv2, PickScore, Aesthetic: ImageReward -0.37 -> -0.19; PickScore 20.07 -> 20.13; Aesthetic 4.85 -> 4.81 | Mixed |
| apply one loop to AlphaFold protein backbone dataset | Ambient Proteins · de novo protein backbone design | diversity vs designability: +14.3% diversity, -0.2% designability | Improved |
Closest papers by meaning
- Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS 2025 · similarity 0.86)
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.85)
- INRFlow: Flow Matching for INRs in Ambient Space (ICML 2025 · similarity 0.85)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.84)
- How Much is a Noisy Image Worth? Data Scaling Laws for Ambient Diffusion. (ICLR 2025 · similarity 0.84)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.84)
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.
