Research map / Image generation / Text-to-image diffusion
Consistent Noisy Latent Rewards for Trajectory Preference Optimization in Diffusion Models
Xiaole Xian, Xilin He, Wenting Chen, Wenshuang Liu, wenqi mu, Yancheng He, Liang Li, Yi Zhang, Xiangyu Yue
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 |
|---|---|---|---|
| add score enhancement to SLRM (ablation) | SLRM · Pick-a-Pic val/test | accuracy: 62.49 -> 67.52 | Improved |
| add task tokens to SLRM (ablation) | SLRM · Pick-a-Pic val/test | accuracy: 65.33 -> 67.52 | Improved |
| use SLRM with score enhancement as TAPO reward model (ablation) | TAPO · Pick-a-Pic v1 val | GenEval: 54.83 -> 64.93 | Improved |
| use task tokens in reward model for TAPO (ablation) | TAPO · Pick-a-Pic v1 val | GenEval: 55.27 -> 64.93 | Improved |
| increase SDE sampling number n from 4 to 8 (ablation) | TAPO · Pick-a-Pic v1 val | GenEval: 56.20 -> 68.93 | Improved |
| increase SDE sampling number n from 8 to 16 (ablation) | TAPO · Pick-a-Pic v1 val | HPSv2.1: 32.01 -> 31.29; 4.52s/iter -> 6.08s/iter | No clear effect |
| increase SDE sampling number n from 2 to 8 (ablation) | TAPO · Pick-a-Pic v1 val | MPS: 2.27 -> 3.07 | Improved |
| diffusion-based SLRM instead of LRM-3.5 | SLRM · Pick-a-Pic val/test (t in [501,1000]) | accuracy: 59.56 -> 66.35 | Improved |
| use SLRM instead of VLM-based HPSv3 on clean images | SLRM · Pick-a-Pic val/test (clean image) | accuracy: 72.80 -> 67.08 | Hurt |
| set beta to 1000 instead of 20 (ablation) | TAPO · Pick-a-Pic v1 val | GenEval: 60.210 -> 68.830 | Improved |
| set beta to 1000 instead of 5000 (ablation) | TAPO · Pick-a-Pic v1 val | GenEval: 63.700 -> 68.830 | Improved |
| TAPO versus LPO | TAPO · Pick-a-Pic v1 val | HPSv3: 9.10 -> 9.89 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- HPSv3 — HPSv3: Towards Wide-Spectrum Human Preference Score (ICCV 2025) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.94)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.91)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.90)
- A Dense Reward View on Aligning Text-to-Image Diffusion with Preference (ICML 2024 · similarity 0.90)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.89)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (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.
