Research map / Image generation / Fast diffusion sampling
Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts
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.
Analyzes the temporal score dynamics of diffusion sampling to find a Profiling/Mutation/Refinement structure, then proposes ASCED, an unsupervised, training-free method that detects artifacts via abnormal score dynamics and corrects them on the fly with trajectory-aware targeted noise injection.
Problem
Visual artifacts persist in diffusion models, and existing detectors are supervised or rely on spatial uncertainty of the final output, which misses when and why artifacts form during generation.
Approach
Artifacts arise in the Mutation phase when pixels enter "score traps" with abnormal score dynamics over time; detect these temporally via thresholded weighted score differences and correct them during sampling by injecting controlled noise only in the flagged regions.
Key results
- ASCED reaches FID 27.66 on LSUN-Horse vs 29.36 for the original DDIM model and beats supervised SARGD and PAL+TTC on FFHQ, LSUN-Horse and LSUN-Bedroom.
- Artifact detection accuracy of the unsupervised method is 56.7% on FFHQ, 67.7% on ImageNet, 65.0% on Bedroom, 68.3% on Cat and 70.3% on Horse, close to supervised PAL but below zero-shot LLaVA on most datasets.
- ASCED runs in about 0.09 s per image, 8.8x faster than PAL (0.79 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ASCED instead of PAL + TTC | Guided Diffusion · FFHQ | FID: 36.35 -> 36.28 | Helped |
| Use ASCED instead of PAL + TTC | Guided Diffusion · ImageNet | FID: 14.01 -> 14.41 | Hurt |
| Use ASCED instead of PAL + TTC | Guided Diffusion · LSUN-Cat | FID: 21.83 -> 21.91 | Hurt |
| Use ASCED instead of PAL + TTC | Guided Diffusion · LSUN-Horse | FID: 28.68 -> 27.66 | Helped |
| Use ASCED instead of PAL + TTC | Guided Diffusion · LSUN-Bedroom | FID: 12.71 -> 12.53 | Helped |
| Use ASCED instead of PAL + TTC | Guided Diffusion · FFHQ | Recall: 0.500 -> 0.503 | Helped |
| Use ASCED instead of PAL + TTC | Guided Diffusion · ImageNet | Recall: 0.747 -> 0.735 | Hurt |
| Use ASCED instead of PAL + TTC | Guided Diffusion · LSUN-Cat | Recall: 0.588 -> 0.593 | Helped |
| Use ASCED instead of PAL + TTC | Guided Diffusion · LSUN-Horse | Recall: 0.646 -> 0.652 | Helped |
| Use ASCED instead of Score Clipping | Guided Diffusion · LSUN-Bedroom | Recall: 0.585 -> 0.590 | Helped |
Surprising findings
- Zero-shot LLaVA-v1.5-13B outperforms the proposed unsupervised detector on most datasets (e.g., 91.1% vs 67.7% on ImageNet), so detection accuracy is not uniformly best for ASCED.
- The paper reports no significant correlation between the number of function evaluations and artifact generation.
Limitations
- Detection struggles on low-contrast images, yielding false negatives.
- Diffusion models that successfully rationalize early abnormal patterns cause false positives.
- Evaluated only on unconditional guided-diffusion models on five datasets with DDIM sampling.
Experimental setup
Datasets: FFHQ ImageNet LSUN-Bedroom LSUN-Cat LSUN-Horse
Benchmarks: FFHQ ImageNet LSUN-Bedroom LSUN-Cat LSUN-Horse
Metrics: FID Precision Recall detection accuracy linear probe accuracy
Baselines: BayesDiff SARGD PAL LLaVA-v1.5-13B State Replacement Score Clipping
Builds on: DDIM SDEdit BayesDiff classifier guidance
Closest papers by meaning
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.90)
- Test-time Scaling of Diffusions with Flow Maps (NeurIPS 2026 · similarity 0.88)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.88)
- Feature Information Dynamics in Diffusion (NeurIPS 2026 · similarity 0.88)
- Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models (CVPR 2024 · similarity 0.88)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · 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.