Research map / Image generation / Fast diffusion sampling
Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models
Cheng Jin, Qitan Shi, Yuantao Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use time-varying guidance (weak early, strong middle, weak late) instead of constant weight | CFG · Text-to-image, NFE=10, guidance 9 | ImageReward: 0.223 -> 0.932 | Improved |
| use time-varying guidance instead of constant weight | CFG · Text-to-image, NFE=10, guidance 9 | diversity: 1.081 -> 1.223 | Improved |
| use time-varying guidance instead of constant weight | CFG · Text-to-image, NFE=10, guidance 9 | saturation: 0.574 -> 0.298 | Improved |
| use time-varying guidance instead of constant weight | CFG · Text-to-image, NFE=10, guidance 9 | FID: 38.988 -> 30.259 | Improved |
| use time-varying guidance at low guidance weight | CFG · Text-to-image, NFE=10, guidance 3 | ImageReward: 0.894 -> 0.859 | No clear effect |
| use time-varying guidance at low NFE | CFG · Text-to-image, NFE=5, guidance 9 | FID: 80.482 -> 46.895 | Improved |
| use TV schedule vs Interval-CFG on FID at high guidance | CFG · Text-to-image, NFE=10, guidance 9 | FID: 28.413 (Interval) vs 30.259 (TV) | Hurt |
| apply time-varying schedule to APG | APG · Text-to-image, NFE=10, guidance 9 | ImageReward: 0.434 -> 0.941 | Improved |
| apply high guidance early and low late instead of early-low (ablation) | CFG · Qualitative bathroom prompt | diversity | Hurt |
| apply high guidance only late (late-high vs constant) (ablation) | CFG · Perturbed-sampling test, futuristic city prompt | fine-grained diversity (MSE): 0.0103 -> 0.0025 | Hurt |
Closest papers by meaning
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.92)
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.91)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.91)
- Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations (NeurIPS 2025 · similarity 0.90)
- Provable Efficiency of Guidance in Diffusion Models for General Data Distribution (ICML 2025 · similarity 0.90)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 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.
