Research map / Image generation / Fast diffusion sampling
Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion 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.
Proposes NCG, a modified classifier-free guidance update that projects the CFG update onto the denoised prediction, down-weights the parallel component, and adds rescaling and reverse momentum to avoid oversaturation at high guidance scales.
Problem
High classifier-free guidance scales improve quality and condition alignment in diffusion models but cause oversaturation and unrealistic artifacts.
Approach
Decompose the CFG update into components parallel and orthogonal to the conditional denoised prediction, down-weight the parallel (saturating) component, and stabilize updates via norm rescaling and negative (reverse) momentum, motivated by viewing CFG as one step of gradient ascent on an l2 distance.
Key results
- On EDM2 class-conditional ImageNet ablation at w=4, NCG reaches FID 6.49, recall 0.62, saturation 0.33, improving over removing projection (FID 6.63) and rescaling (FID 7.93).
- NCG improves FID, recall, and saturation over CFG across multiple models while keeping precision comparable to CFG.
- NCG adds negligible overhead: about 0.45 ms guidance step versus about 130 ms forward pass for SDXL on an RTX 3090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CFG with NCG | Stable Diffusion · Stable Diffusion CLIP-Score | CLIP-Score: 0.31 -> 0.31 | No clear effect |
| Replace CFG with NCG | ImageNet generation · ImageNet class accuracy | class accuracy: 0.97 -> 0.96 | Mixed |
| Add projection (ablation) | NCG · ImageNet class-conditional generation (EDM2, w=4) | FID: 6.63 -> 6.49 | Helped |
| Add projection (ablation) | NCG · ImageNet class-conditional generation (EDM2, w=4) | Saturation: 0.37 -> 0.33 | Helped |
| Add rescaling (ablation) | NCG · ImageNet class-conditional generation (EDM2, w=4) | FID: 7.93 -> 6.49 | Helped |
| Add rescaling (ablation) | NCG · ImageNet class-conditional generation (EDM2, w=4) | Recall: 0.56 -> 0.62 | Helped |
| Add reverse momentum (ablation) | NCG · ImageNet class-conditional generation (EDM2, w=4) | FID: 6.85 -> 6.49 | Helped |
| Add reverse momentum (ablation) | NCG · ImageNet class-conditional generation (EDM2, w=4) | Recall: 0.61 -> 0.62 | Helped |
| Increase the parallel-component weight η (ablation) | NCG · ImageNet class-conditional generation | Saturation | Hurt |
| Increase the parallel-component weight η (ablation) | NCG · ImageNet class-conditional generation | FID | Hurt |
| Use an excessively small rescaling radius r (ablation) | NCG · ImageNet class-conditional generation | image quality | Hurt |
| Use a high rescaling radius r (ablation) | NCG · ImageNet class-conditional generation | image quality | No clear effect |
Surprising findings
- Projecting the CFG update onto the predicted noise gives results nearly identical to standard CFG; projection must be applied to the denoised prediction to be effective.
Limitations
- Still requires two denoiser queries per step like CFG, so sampling cost is not reduced to unguided levels.
- Hyperparameters (eta, rescaling radius r, momentum beta) need tuning per setup.
Experimental setup
Datasets: ImageNet MS COCO 2017
Benchmarks: ImageNet MS COCO 2017
Metrics: FID precision recall saturation contrast class accuracy CLIP score
Baselines: CFG CFG Rescale CADS interval guidance dynamic thresholding ICG TSG
Builds on: classifier-free guidance EDM normalized gradient ascent CFG Rescale
Closest papers by meaning
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.93)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.93)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.93)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.93)
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.92)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.92)
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.