Research map / Image generation / Fast diffusion sampling
Ambient Diffusion Omni: Training Good Models with Bad Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Ambient Diffusion Omni, which trains diffusion models on all available data (low-quality, synthetic, out-of-distribution) by annotating each sample with a noise level above which it is usable (high noise) and using locality-based crop annotations for low noise, achieving SOTA ImageNet FID and better diversity in text-to-image.
Problem
Standard diffusion training discards low-quality, synthetic, and out-of-distribution images via filtering, which reduces data diversity; the goal is to extract useful signal from heterogeneous-quality data without knowing the degradation process.
Approach
Use a time-conditional classifier to find, per sample, the minimal noise level at which a corrupted sample is indistinguishable from clean data, and train only on it for diffusion times above that level; at low noise, use crops-classifier-based annotations so out-of-distribution images contribute local patches, exploiting spectral power law and locality of natural images.
Key results
- On ImageNet-512 with XXL model, Ambient-o+crops reaches test FID 2.53 with CFG versus 2.73 for EDM2-XXL.
- In the controlled CIFAR-10 setting with 10% clean data and 90% Gaussian-blurred data, classifier annotations reduce FID from 8.79 (clean only) to 5.34 at blur sigma 0.6.
- On text-to-image, Ambient-o achieves COCO FID 10.61 versus 12.37 for the MicroDiffusion baseline, with DINO Vendi diversity 3.65 vs 3.22 compared to fine-tuning on the high-quality subset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ambient-o training instead of standard training | EDM2-XS · ImageNet-512 | Test FID: 3.77 -> 3.69 | Helped |
| Use Ambient-o training instead of standard training | EDM2-XXL · ImageNet-512 | Test FID: 2.88 -> 2.81 | Helped |
| Add crop-based training (ablation) | Ambient-o-XXL · ImageNet-512 | Test FID: 2.81 -> 2.78 | Helped |
| Use Ambient-o annotations rather than training on all data equally | CIFAR-10 diffusion model · CIFAR-10, Gaussian blur σB=1.0 | FID: 45.32 -> 6.16 | Helped |
| Use Ambient-o annotations rather than training on all data equally | CIFAR-10 diffusion model · CIFAR-10, Gaussian blur σB=0.8 | FID: 28.26 -> 6.00 | Helped |
| Use Ambient-o annotations rather than training on all data equally | CIFAR-10 diffusion model · CIFAR-10, Gaussian blur σB=0.6 | FID: 11.42 -> 5.34 | Helped |
| Use Ambient-o annotations rather than training on all data equally | CIFAR-10 diffusion model · CIFAR-10, Gaussian blur σB=0.4 | FID: 2.47 -> 2.44 | No clear effect |
| Use classifier-based crop annotation instead of fixed σ=0.1 (ablation) | Dogs diffusion model · CIFAR-10, dogs from cats | FID: 9.85 -> 8.92 | Helped |
| Use high-quality-subset fine-tuning instead of Ambient-o (ablation) | MicroDiffusion · COCO | FID: 10.61 vs 10.27 | Helped |
| Use Ambient-o with all data instead of high-quality-subset fine-tuning (ablation) | MicroDiffusion · DiffDB prompts | DINO Vendi Diversity: 3.22 -> 3.65; >13% | Helped |
| Add the ambient preconditioning weight (ablation) | CIFAR-10 Ambient-o · CIFAR-10, Gaussian blur σB=0.6 | FID: 5.49 -> 5.36 | Helped |
| Clip the ambient weight at 4.0 rather than using the unclipped weight (ablation) | CIFAR-10 Ambient-o · CIFAR-10, Gaussian blur σB=0.6 | FID: 5.36 -> 5.69 | Hurt |
Surprising findings
- Training on fully blurred or corrupted data alone (no annotation) yields very poor FID (e.g., 45.32 at sigma 1.0), but adding noise-level annotations recovers and surpasses clean-only performance.
- Some ImageNet images with low global quality contain high-quality crops that are useful for low-noise denoising.
Limitations
- Full method requires training annotation classifiers, increasing runtime.
- Less effective for low-frequency corruptions such as color shifts, contrast reduction, or masking.
- Assumes known labeled clean and corrupted subsets or a quality metric to construct them.
Experimental setup
Datasets: CIFAR-10 FFHQ AFHQ ImageNet Conceptual Captions Segment Anything JourneyDB DiffusionDB
Baselines: EDM-2 MicroDiffusion EDM Only clean data filtering Fine-tuning on high-quality subset
Builds on: Ambient Diffusion EDM EDM2 MicroDiffusion classifier guidance
Closest papers by meaning
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.91)
- Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models (ICLR 2025 · similarity 0.90)
- Stochastic Forward–Backward Deconvolution: Training Diffusion Models with Finite Noisy Datasets (ICML 2025 · similarity 0.89)
- One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls (CVPR 2024 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.89)
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.