Research map / Image generation / Fast diffusion sampling
Applying Guidance in a Limited Interval Improves Sample and Distribution Quality 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.
Limits classifier-free guidance to a middle interval of noise levels (sigma_lo, sigma_hi] instead of all steps, improving FID on ImageNet-512 with EDM2 (1.81 to 1.40) and sample quality/speed on SD-XL.
Problem
Classifier-free guidance is applied with a constant weight across all sampling steps, but it is harmful at high noise levels (reducing diversity, mode dropping) and largely unnecessary at low noise levels, so constant guidance is suboptimal for image quality and sampling cost.
Approach
Apply guidance only within a continuous interval of noise levels in the middle of the sampling chain (disabling it elsewhere, with the weight set to 1), exposing the guidance interval as a sampler hyperparameter; the interval endpoints are aligned to step boundaries so the ODE remains smooth within each step.
Key results
- On ImageNet-512 with EDM2-XXL, limited guidance interval improves FID from 1.81 to 1.40 (record), and FD_DINOv2 from 33.09 to 29.16.
- With EDM2-S, FID improves from 2.23 to 1.68.
- On SD-XL, guidance applied in 50% of steps (sigma in (0.28, 5.42], w=16) yields over 20% speed-up due to fewer unconditional evaluations, with visually improved diversity and reduced oversaturation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| restrict guidance to a middle noise-level interval | EDM2-S · ImageNet-512 | FID: 2.23 to 1.68 | Helped |
| restrict guidance to a middle noise-level interval | EDM2-XXL · ImageNet-512 | FID: 1.81 to 1.40 | Helped |
| restrict guidance to a middle noise-level interval | EDM2-XXL · ImageNet-512 | FD-DINO: 33.09 to 29.16 | Helped |
| restrict guidance to a middle noise-level interval | EDM2-S · ImageNet, 16 sampling steps | FID: 2.49 to 1.84 | Helped |
| restrict guidance to a middle noise-level interval | EDM2-S · ImageNet, 64 sampling steps | FID: 2.27 to 1.70 | Helped |
| apply guidance at 50% of sampling steps | Stable Diffusion XL · Stable Diffusion XL sampling | speed-up: over 20% | Helped |
| include higher noise levels in the guidance interval (ablation) | limited guidance interval · ImageNet | FID | Hurt |
| narrow the guidance interval by lowering its upper limit (ablation) | limited guidance interval · ImageNet | FID | Hurt |
| apply guidance at low noise levels (ablation) | limited guidance interval · ImageNet | FID | No clear effect |
| replace binary interval guidance with smooth weighting functions (ablation) | limited guidance interval · ImageNet | FID | No clear effect |
| estimate guidance importance by toggling it at one sampling step at a time (ablation) | limited guidance interval · ImageNet | guidance-interval sensitivity | Hurt |
| disable guidance at high noise levels (ablation) | Stable Diffusion XL · Stable Diffusion XL samples | image composition and color saturation | Mixed |
Surprising findings
- Standard CFG is harmful at high noise levels and disabling it there improves results, while disabling it at low noise levels has little effect.
- Smooth weighting functions over the guidance interval did not improve over simple binary inclusion.
Limitations
- Optimal interval and weight are found by grid search per model, and for SD-XL chosen by visual inspection.
- Assumes deterministic Runge-Kutta-style samplers; interval endpoints must align to step boundaries.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 512x512
Metrics: FID FD_DINOv2 precision recall
Baselines: EDM2-S EDM2-XXL DiT-XL/2 SD-XL standard classifier-free guidance dynamic CFG
Builds on: classifier-free guidance EDM EDM2 Heun sampler DiT
Closest papers by meaning
- Provable Efficiency of Guidance in Diffusion Models for General Data Distribution (ICML 2025 · similarity 0.91)
- What does guidance do? A fine-grained analysis in a simple setting (NeurIPS 2024 · similarity 0.91)
- Theoretical insights for diffusion guidance: A case study for Gaussian mixture models (ICML 2024 · similarity 0.91)
- REG: Rectified Gradient Guidance for Conditional Diffusion Models (ICML 2025 · similarity 0.91)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.90)
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.