Research map / Image generation / Text-to-image diffusion
ViPO: Visual Preference Optimization at Scale
Ming Li, Jie Wu, Justin Cui, Xiaojie Li, Rui Wang, Chen Chen
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 polynomial confidence term (alpha=8) to Diffusion-DPO | Diffusion-DPO SD1.5 · GenEval (Pick-a-Pic V2 training) | overall: 43.00 -> 49.87 | Improved |
| add polynomial confidence term to Diffusion-DPO | Diffusion-DPO SDXL · GenEval (Pick-a-Pic V2 training) | overall: 58.02 -> 60.34 | Improved |
| add polynomial confidence term to Diffusion-DPO | Diffusion-DPO SD1.5 · Pick-a-Pic V2 test | ImageReward: 0.297 -> 0.679 | Improved |
| add polynomial confidence term to Diffusion-DPO | Diffusion-DPO SD1.5 · Pick-a-Pic V2 test | HPSv2.1: 26.12 -> 28.30 | Improved |
| add polynomial confidence term vs Diffusion-KTO | Diffusion-KTO SD1.5 · Pick-a-Pic V2 test | PickScore: 21.06 -> 21.48 | Improved |
| use negative alpha on over-simple shuffled-loser dataset (ablation) | Poly-DPO SD1.5 · synthetic over-simple preference dataset | HPSv2.1 (figure) | Improved |
| vary alpha on high-quality ViPO-Image-1M (ablation) | Poly-DPO SD1.5 (SFT-initialized) · ViPO-Image-1M | HPSv2.1 (figure) | No clear effect |
| add SFT then Poly-DPO on ViPO-Image-1M | SD1.5 · GenEval | overall: 0.42 -> 0.54 | Improved |
| add Poly-DPO after SFT on ViPO-Image-1M | SD3.5-Medium SFT · DPG-Bench | overall: 84.24 -> 87.71 | Improved |
| add Poly-DPO after SFT on ViPO-Image-1M | FLUX.1-dev SFT · CVTG-2K | word accuracy: 0.2126 -> 0.6859 | Improved |
| add Poly-DPO after SFT on ViPO-Image-1M | SD3.5-Medium SFT · GPT-4o human quality | accuracy: 77.50 -> 85.25 | Improved |
| add SFT stage before Poly-DPO (ablation) | Poly-DPO SD1.5 · SD1.5 SFT ablation | ImageReward: 0.6391 -> 1.0161 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SDXL base — SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (ICLR 2024) · 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
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.90)
- DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025 · similarity 0.89)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.89)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.89)
- Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization (CVPR 2025 · similarity 0.88)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.87)
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.
