Research map / Image generation / Fast diffusion sampling
Fine-Tuning Diffusion Models via Intermediate Distribution Shaping
Gautham Govind Anil, Shaan Ul Haque, Nithish Kannen, Dheeraj Mysore Nagaraj, Sanjay Shakkottai, Karthikeyan Shanmugam
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 GRAFT rejection-sampling fine-tuning instead of DDPO | SD v2 · GenAI-Bench (fine-tuned on GenAI-Bench) | VQAScore: 65.70 -> 70.51 | Improved |
| train DDPO policy gradient | SD v2 · GenAI-Bench | VQAScore: 66.87 -> 65.70 | No clear effect |
| fine-tune only the early denoising steps (P-GRAFT 0.25N) instead of all steps (GRAFT) | GRAFT · GenAI-Bench | VQAScore: 70.51 -> 71.94 | Improved |
| fine-tune only the early denoising steps (P-GRAFT 0.25N) instead of all steps (GRAFT) | GRAFT · GenEval (fine-tuned on T2ICompBench++) | VQAScore: 79.57 -> 80.29 | Improved |
| switch point N_I = 0.75N instead of 0.25N (ablation) | P-GRAFT · GenAI-Bench | VQAScore: 71.94 -> 69.46 | Hurt |
| switch point N_I = 0.5N instead of 0.25N (ablation) | P-GRAFT · GenEval (fine-tuned on GenAI-Bench) | VQAScore: 80.96 -> 80.60 | Hurt |
| reverse P-GRAFT (fine-tuned model in later steps) (ablation) | P-GRAFT · SD v2 text-to-image | VQAScore | Hurt |
| vary LoRA rank (ablation) | P-GRAFT · SD v2 text-to-image | VQAScore | No clear effect |
| P-GRAFT 0.25N instead of GRAFT on layouts | GRAFT · PubLayNet unconditional | alignment (lower better): 0.064 -> 0.053 | Improved |
| P-GRAFT 0.25N instead of GRAFT on layouts | GRAFT · PubLayNet unconditional | FID: 10.68 -> 9.91 | Hurt |
| P-GRAFT 0.25N instead of GRAFT on molecules | GRAFT · QM9 | molecule stability: 90.76 -> 92.61 | Improved |
| P-GRAFT 0.5N instead of baseline on molecules | baseline · QM9 | molecule stability: 90.50 -> 90.46 | No clear effect |
Closest papers by meaning
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.89)
- Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization (ICLR 2025 · similarity 0.89)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.89)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · 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.
