Research map / Image generation / Fast diffusion sampling
Not All Prediction Targets Keep Training-Free Diffusion Guidance on the Manifold
Yunsung Lee, Hyeongmin Lee
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 x-prediction instead of eps-prediction (ablation) | DPS-guided flow model · Crossed-lines toy, D=512 | on-manifold rate (%): 0.5 -> 93.3 | Improved |
| use v-prediction instead of eps-prediction (ablation) | DPS-guided flow model · Crossed-lines toy, D=512 | on-manifold rate (%): 0.5 -> 21.5 | Improved |
| use x-prediction (JiT-H) instead of eps-prediction (DiT) (ablation) | DPS guidance · Bird benchmark, matched Validity 26.6% | C-FID: 38.1 -> 32.9 | Improved |
| use x-prediction (JiT-H) instead of v-prediction (SiT) (ablation) | DPS guidance · Bird benchmark, matched Validity 26.6% | C-FID: 34.7 -> 32.9 | Improved |
| strengthen DPS guidance on eps-prediction (rho 0.1 to 0.5) | DiT · Bird benchmark | P-FID: 6.7 -> 14.2 | Hurt |
| strengthen guidance to rho=8 on x-prediction | JiT-H · Bird benchmark | C-FID: 34.5 -> 30.6 | Improved |
| use x-prediction instead of v-prediction in pixel space under strong guidance (ablation) | DPS guidance · Bird benchmark | C-FID: 47.7 -> 30.6 | Improved |
| increase JiT capacity from B to H (ablation) | JiT · Bird benchmark | C-FID: 31.3 -> 30.6 | Improved |
| use x-prediction instead of eps-prediction (ablation) | DPS guidance · Bird benchmark | Recall (DINOv2): 0.49 -> 0.59 | Improved |
| raise DPS strength from rho=0 to rho=10 | DiT · Style transfer | Content Accuracy (%): 89 -> 1.5 | Hurt |
| use x-prediction instead of eps-prediction at ~80% content accuracy (ablation) | DPS style guidance · Style transfer | Gram Distance: 5.62 -> 5.48 | Improved |
| use x-prediction instead of v-prediction (SiT) at rho=4 (ablation) | DPS style guidance · Style transfer | Content Accuracy (%): 71 -> 87 | Improved |
Closest papers by meaning
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.89)
- Diffusion Models without Classifier-free Guidance (NeurIPS 2026 · similarity 0.89)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.89)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.89)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.89)
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.
