Research map / Image generation / Fast diffusion sampling
RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models
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.
Reinterprets diffusion sampling as a continuation method and proposes RODS, a plug-and-play sampler that detects high-curvature score-field steps and applies robust (min-max) corrective updates to reduce hallucinations in EDM-based image generation.
Problem
Diffusion samplers can produce hallucinated outputs (e.g., malformed faces, hands) due to inaccurate score approximation in low-density regions, and there is no standard way to detect or correct such failures.
Approach
Views each diffusion step as a proximal/gradient update on f_t = -log p_t in a continuation framework, detects risky steps via a curvature-based index on the score-norm field, and applies robust optimization inspired updates (SAS or CAS) only at flagged steps, without retraining.
Key results
- On AFHQv2 (1,080 samples, epsilon=0.1), RODS-CAS detects 87.5% of labeled hallucinations and corrects 10 cases with no degraded cases.
- On 11k-hands, RODS-CAS reduces hallucination rate from 19.4% (EDM-Euler) to 14.3% with 26.3% correction and 0.0% new hallucinations, at 4.82 s vs 2.71 s inference time.
- On FFHQ, RODS-CAS detects 72.5% of labeled hallucinations, while FID rises from 20.48 (EDM-Euler) to 22.73.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RODS-CAS instead of EDM-Euler | EDM sampling · 11k-hands | Hallucination rate: 19.4% -> 14.3% | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · 11k-hands | Correction rate: +26.3% | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · 11k-hands | New hallucination rate: 0.0% | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · 11k-hands | Inference time: 2.71 -> 4.82 | Hurt |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · AFHQv2 | Hallucination detection rate: 87.5% | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · AFHQv2 | Hallucination correction count: 10 cases | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · FFHQ | Hallucination detection rate: 72.5% | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · FFHQ | Hallucination correction count: 9 cases | Helped |
| Use RODS-CAS instead of EDM-Euler | EDM sampling · 11k-hands | Hallucination detection count: 169 out of 175 | Helped |
| Increase Euler sampling from 40 to 200 steps (ablation) | EDM-Euler · 11k-hands | Hallucination rate: 19.4% -> 16.9% | Helped |
| Use 5-step or 10-step optimization instead of the single-step update (ablation) | RODS · AFHQ | Perceptible output difference count: 3 out of 184 | No clear effect |
| Use 5-step or 10-step optimization instead of the single-step update (ablation) | RODS · FFHQ | Perceptible output difference count: 4 out of 222 | No clear effect |
Surprising findings
- Simply increasing EDM-Euler to 200 steps does not remove the hallucinations, which the authors attribute to directional error rather than step size.
- On FFHQ and 11k-hands, RODS-CAS has noticeably worse FID than EDM-Euler (e.g., 16.68 vs 14.57 on 11k-hands), despite claims of comparable quality.
Limitations
- Hallucination labels are produced by a small number of lab members' manual annotation with no standard benchmark, limiting reproducibility and generality.
- Experiments are limited to unconditional VE-ODE EDM sampling; extension to VP, latent diffusion, and text-to-image is not validated.
- Detection threshold and perturbation radius must be tuned per dataset, and the method flags many false positives at low thresholds.
Experimental setup
Datasets: AFHQv2 FFHQ 11k-hands
Benchmarks: AFHQv2 FFHQ 11k-hands
Metrics: hallucination rate correction rate new hallucination rate FID Inception variance DreamSim diversity inference time true positive rate false positive rate
Baselines: EDM-Euler EDM-Heun EDM-Euler-200
Builds on: EDM continuation method robust optimization sharpness-aware minimization
Closest papers by meaning
- Understanding Hallucinations in Diffusion Models through Mode Interpolation (NeurIPS 2024 · similarity 0.90)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.88)
- Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · similarity 0.88)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.87)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.87)
- Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization (NeurIPS 2024 · similarity 0.87)
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.