Research map / Image generation / Fast diffusion sampling
Beyond VLM-Based Rewards: Diffusion-Native Latent Reward Modeling
Gongye Liu, Bo Yang, Zhi Yida, Zhizhou Zhong, Lei Ke, Didan Deng, Han Gao, Yongxiang Huang, Kaihao Zhang, Hongbo Fu, Wenhan Luo
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 multi-noise test-time ensembling (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 71.49 -> 72.48 | Improved |
| use noise-calibrated variance instead of fixed variance (ablation) | DiNa-LRM · HPDv2 | pairwise preference accuracy (%): 78.72 -> 82.13 | Improved |
| use noise-calibrated variance with ensembling (ablation) | DiNa-LRM* · Average (4 sets) | pairwise preference accuracy (%): 70.41 -> 72.48 | Improved |
| train with uniform timestep sampling instead of constant t=0 (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 64.93 -> 70.68 | Improved |
| train with uniform timestep sampling instead of LogitNormal(0.8,1.0) (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 70.58 -> 71.49 | Improved |
| LoRA fine-tune backbone instead of freezing (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 70.27 -> 71.49 | Improved |
| use gated Q-Former head instead of average pooling (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 71.03 -> 71.49 | Improved |
| use gated Q-Former head instead of transformer plus average pooling (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 71.14 -> 71.49 | Improved |
| increase feature layers from 8 to 12 (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 68.07 -> 71.49 | Improved |
| increase feature layers from 12 to 20 (ablation) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 71.49 -> 72.76 | Improved |
| add multi-noise ensembling to fixed-variance model (ablation) | DiNa-LRM (Fixed variance) · Average (4 sets) | pairwise preference accuracy (%): 70.68 -> 70.41 | Hurt |
| scale diffusion backbone from SD3.5-M (2B) to FLUX.1-Dev (12B) | DiNa-LRM · Average (4 sets) | pairwise preference accuracy (%): 71.49 -> 69.87 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- HPSv3 — HPSv3: Towards Wide-Spectrum Human Preference Score (ICCV 2025) · better on 0 of 5 shared results
- ImageReward — ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023) · better on 1 of 1 shared result
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 0.93)
- PRDP: Proximal Reward Difference Prediction for Large-Scale Reward Finetuning of Diffusion Models (CVPR 2024 · similarity 0.90)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.90)
- RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · similarity 0.89)
- Preference Learning with Response Time: Robust Losses and Guarantees (NeurIPS 2025 · similarity 0.88)
- Dynamics-Aware Preference Optimization for Vision-Language Models (CVPR 2026 · 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.
