Research map / Vision-language and multimodal / Multimodal LLMs
DeCoRL: Decomposed Consistency Reinforcement Learning for Multi-Image Composition
Zhiqiang Wu, Shuang Sun, Jiale Zhang, Jing Li, Yibo Chen, Xu Tang, Yao Hu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeCoRL, a decomposed consistency reinforcement learning approach for multi-image composition (MICo) in image editing models, motivated by SFT's weak object-level correspondence supervision.
Problem
Multi-image composition in large-scale image editing models, where consistency degrades as the number of reference images grows; SFT gives only implicit supervision on the composed image and fails to learn reliable object-level correspondence across references.
Approach
Decomposes consistency into explicit object-level correspondence signals and optimizes them with reinforcement learning rather than relying on implicit image-level supervision from SFT. Abstract is truncated, so the exact reward design is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DeCoRL object-level RL after SFT (ablation) | Qwen-Image-Edit2511 + SFT · DeCo-Bench (overall) | consistency score: 7.05 -> 8.16 | Improved |
| add DeCoRL after SFT, 6 reference images (ablation) | Qwen-Image-Edit2511 + SFT · DeCo-Bench (6 images) | consistency score: 6.15 -> 7.61 | Improved |
| add multi-image SFT before RL (ablation) | Qwen-Image-Edit2511 · DeCo-Bench (overall) | consistency score: 6.22 -> 7.05 | Improved |
| add multi-image SFT, 2 reference images (ablation) | Qwen-Image-Edit2511 · DeCo-Bench (2 images) | consistency score: 8.41 -> 8.34 | No clear effect |
| add DeCoRL after SFT (ablation) | Qwen-Image-Edit2511 + SFT · MICo-Bench (overall) | overall score: 22.90 -> 37.58 -> 54.03 | Improved |
| add DeCoRL after SFT (ablation) | Qwen-Image-Edit2511 + SFT · MICo-Bench (Naturalness) | Naturalness: 6.54 -> 8.01 | Improved |
| add DeCoRL after SFT (ablation) | Qwen-Image-Edit2511 + SFT · MICo-Bench (Prompt Following) | PF: 8.76 -> 8.69 | No clear effect |
| add DeCoRL after SFT (ablation) | Qwen-Image-Edit2511 + SFT · MICo-Bench (Subject Resemblance) | SR: 7.01 -> 7.64 | Improved |
| use a CoT-trained single-object reward VLM instead of the vanilla VLM (ablation) | DeCoRL · DeCo-Bench (overall) | consistency score: 7.21 -> 8.16 | Improved |
| use a vanilla VLM as the reward in place of no RL (ablation) | Qwen-Image-Edit2511 + SFT · DeCo-Bench (4 images) | consistency score: 6.80 -> 6.42 | Hurt |
| fine-tune the reward VLM with chain-of-thought supervision (ablation) | Qwen3-VL-32B reward model · EditReward-Bench | human-label agreement accuracy: 0.354 -> 0.721 | Improved |
| fine-tune the reward VLM with chain-of-thought supervision (ablation) | DeCoRL training dynamics · multi-object consistency score during RL | reward level: below 0.5 (vanilla) vs above 0.6 (trained) | Improved |
Experimental setup
Baselines: supervised fine-tuning
Closest papers by meaning
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.85)
- MICON-Bench: Benchmarking and Enhancing Multi-Image Context Image Generation in Unified Multimodal Models (CVPR 2026 · similarity 0.85)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.84)
- Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models (NeurIPS 2026 · similarity 0.84)
- EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.84)
- The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.84)
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.
