Research map / Image generation / Text-to-image diffusion
PaCo-RL: Advancing Reinforcement Learning for Consistent Image Generation with Pairwise Reward Modeling
Bowen Ping, Chengyou Jia, Minnan Luo, Changliang Xia, Xin Shen, Zhuohang Dang, Hangwei Qian
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 resolution-decoupled training (0.5x) (ablation) | PaCo-GRPO FLUX.1-dev · T2IS ablation | Prompt following: 0.698 -> 0.728 | Improved |
| add resolution-decoupled training (0.5x) (ablation) | PaCo-GRPO FLUX.1-dev · T2IS ablation | Training time: 12.0h -> 6.0h | Improved |
| add resolution-decoupled training (ablation) | PaCo-GRPO FLUX.1-dev · T2IS ablation | Visual consistency: 0.452 -> 0.493 | Improved |
| add log-tamed reward aggregation (ablation) | PaCo-GRPO FLUX.1-dev · T2IS ablation | Aesthetics: 0.471 -> 0.555 | Improved |
| add log-tamed reward aggregation (ablation) | PaCo-GRPO FLUX.1-dev · T2IS ablation | Visual consistency: 0.557 -> 0.493 | Hurt |
| add log-tamed reward aggregation (ablation) | PaCo-GRPO FLUX.1-dev · reward ratio during training | Consistency/Prompt Alignment ratio: above 2.5 -> below 1.8 | Improved |
| train at 256x256 instead of 512x512 (ablation) | PaCo-GRPO FLUX.1-dev · T2IS training | reward | Hurt |
| train at 512x512 instead of 1024x1024 (ablation) | PaCo-GRPO FLUX.1-dev · T2IS training | reward | No clear effect |
| fine-tune Qwen2.5-VL-7B on PaCo-Dataset into PaCo-Reward-7B | Qwen2.5-VL-7B · ConsistencyRank | Accuracy: 0.344 -> 0.449 | Improved |
| fine-tune Qwen2.5-VL-7B on PaCo-Dataset | Qwen2.5-VL-7B · EditReward-Bench | Overall accuracy: 0.432 -> 0.751 | Improved |
| train with reasoning-augmented labels instead of binary labels only | PaCo-Reward-7B · ConsistencyRank | Accuracy: 0.441 -> 0.449 | Improved |
| add PaCo-Reward-7B RL training | FLUX.1-Kontext · GEdit-Bench EN-I | Overall: 6.143 -> 6.636 | 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.
- DreamSim — DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data (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
- Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · similarity 0.88)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Auto-Rubric as Reward: From Implicit Preference to Explicit Generative Criteria (NeurIPS 2026 · similarity 0.87)
- Generalizing Consistency Policy to Visual RL with Prioritized Proximal Experience Regularization (NeurIPS 2024 · similarity 0.86)
- MSRL: Scaling Generative Multimodal Reward Modeling via Multi-Stage Reinforcement Learning (CVPR 2026 · similarity 0.86)
- PrefPaint: Aligning Image Inpainting Diffusion Model with Human Preference (NeurIPS 2024 · similarity 0.86)
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.
