Research map / Image generation / Text-to-image diffusion
D2PO: Optimizing Diffusion Samplers via Dynamic Preference
Jinkyu Kim, Jinyoung Choi, Bohyung Han
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 |
|---|---|---|---|
| add dynamic preference (2N-step self-refined teacher instead of fixed N+1-step teacher) (ablation) | D2PO · SD v1.5 COCO iPNDM, 4 steps | FID: 16.91 -> 15.69 | Improved |
| add dynamic preference (ablation) | D2PO · SD v1.5 COCO iPNDM, 5 steps | FID: 13.70 -> 13.38 | Improved |
| use score-based energy instead of LPIPS (ablation) | D2PO · SD v1.5 COCO iPNDM, 4 steps | FID: 17.88 -> 15.69 | Improved |
| use score-based energy instead of LPIPS (ablation) | D2PO · SD v1.5 COCO iPNDM, 5 steps | FID: 14.59 -> 13.38 | Improved |
| use score-based energy instead of LPIPS (ablation) | D2PO · SD v1.5 COCO iPNDM, 4 steps | Aesthetic score: 5.1796 -> 5.2024 | Improved |
| use epoch-wise copy instead of EMA for reference timestep schedule (ablation) | D2PO · SD v1.5 COCO iPNDM, 4 steps | FID: 15.75 -> 15.69 | Improved |
| use epoch-wise copy instead of EMA for reference timestep schedule (ablation) | D2PO · SD v1.5 COCO iPNDM, 4 steps | Aesthetic score: 5.1937 -> 5.2024 | Improved |
| D2PO instead of LD3 sampler optimization | iPNDM sampler · SD v1.5 COCO iPNDM, 4 steps | FID: 17.60 -> 15.69 | Improved |
| D2PO instead of LD3 at 7 steps | iPNDM sampler · SD v1.5 COCO iPNDM, 7 steps | FID: 12.41 -> 12.71 | Hurt |
| D2PO instead of LD3 | iPNDM sampler · SD v1.5 COCO iPNDM, 7 steps | HPSv2: 0.2434 -> 0.2513 | Improved |
| D2PO instead of LD3 | iPNDM 3M sampler · ImageNet-256 latent, 4 steps | FID: 9.19 -> 7.28 | Improved |
| D2PO instead of LD3 | InstaFlow sampler · InstaFlow COCO, 2 steps | FID: 63.55 -> 40.68 | Improved |
Closest papers by meaning
- Inference-Time Alignment of Diffusion Models with Direct Noise Optimization (ICML 2025 · similarity 0.90)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.89)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.89)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.89)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.89)
- Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps (ICLR 2024 · similarity 0.88)
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.
