Research map / Reinforcement learning / RL from feedback
Threshold-Guided Optimization for Visual Generative Models
Jinbin Bai, Yu Lei, Qingyu Shi, Aosong Feng, Yi Xin, Zhuoran Zhao, Fei Shen, Kaidong Yu, Xiangtai Li
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 TGO threshold-guided loss instead of SFT on pseudo-preferred samples | SD v1.5 · Pick-a-Pic | ImageReward: 0.5015 -> 0.6703 | Improved |
| use TGO instead of Diffusion-KTO | SD v1.5 · HPSv2 | HPSv2.1: 0.2873 -> 0.2961 | Improved |
| use TGO instead of Diffusion-DPO | SD v1.5 · Pick-a-Pic | PickScore: 20.78 -> 21.05 | Improved |
| apply TGO on scalar-feedback data vs SFT | SD v1.4 · HPS Prompts | ImageReward mean: 0.2348 -> 0.3523 | Improved |
| apply TGO vs SFT on masked generative model | Meissonic · HPS Prompts | HPSv2.1 mean: 0.2912 -> 0.2915 | No clear effect |
| apply TGO vs SFT on PickScore | SD v1.4 · HPS Prompts | PickScore mean: 20.7217 -> 20.9001 | Improved |
| use TGO-LoRA instead of KTO-LoRA for video | Wan 1.3B · VideoAlign | overall score: -1.3703 -> -1.1757 | Improved |
| use TGO-LoRA instead of SFT-LoRA for temporal alignment | Wan 1.3B · VideoAlign | TA score: -0.6705 -> -0.6753 | No clear effect |
| lower threshold percentile from 0.5 to 0.1 (ablation) | TGO · Pick-a-Pic | ImageReward: 0.6728 -> 0.4913 | Hurt |
| raise threshold percentile from 0.5 to 0.9 (ablation) | TGO · Pick-a-Pic | ImageReward: 0.6728 -> 0.3385 | Hurt |
| use p=0.3 instead of median (ablation) | TGO · Pick-a-Pic, PartiPrompts, HPSv2 | ImageReward: 0.6728 -> 0.6817 | No clear effect |
Closest papers by meaning
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.91)
- RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · similarity 0.89)
- Aligning Few-Step Generative Model via Amortizing Sample-Based Variational Inference (NeurIPS 2026 · similarity 0.89)
- K-Sort Arena: Efficient and Reliable Benchmarking for Generative Models via K-wise Human Preferences (CVPR 2025 · similarity 0.89)
- Aligning Visual Contrastive learning models via Preference Optimization (ICLR 2025 · similarity 0.88)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 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.
