Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing

2025-06-22 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

Edit-R2 is a novel reinforcement learning framework designed for multi-turn in-context image editing, addressing challenges like long-context dilution and state contamination. Developed by Hong Kong University of Science and Technology and Kuaishou Technology, it reconstructs operative session intent using in-context chain-of-thought (IC-CoT) and employs a unified multi-turn RL objective for both discrete text reasoning and continuous latent space image generation. A trajectory filtering mechanism stabilizes training. To support evaluation, Edit-R2 introduces MICE-Bench, a large-scale benchmark with 720 three-turn editing instances, using automated metrics for instruction following (IF), content consistency (CC), and global awareness (GA). Experiments show Edit-R2 significantly improves multi-turn editing, achieving gains of +18% IF and +18% GA over BAGEL at Turn 2, and +8% IF and +15% GA at Turn 3, outperforming other open-source models and competing with closed-source counterparts.

Key takeaway

For AI Scientists and Machine Learning Engineers developing interactive image editing systems, traditional single-turn models are inadequate for realistic multi-turn user sessions. You should consider integrating explicit session intent reconstruction, like Edit-R2's IC-CoT, and unified multi-turn reinforcement learning. This approach effectively mitigates long-context dilution and error propagation, crucial for maintaining consistency and instruction adherence across iterative edits, as demonstrated by Edit-R2's significant performance gains on MICE-Bench.

Key insights

Multi-turn image editing benefits from explicit session intent reconstruction and unified RL to manage long-context and error propagation.

Principles

• Explicitly reconstruct session intent from interleaved image-text history.
• Jointly optimize discrete reasoning and continuous visual generation.
• Filter corrupted rollouts to stabilize multi-turn reinforcement learning.

Method

Edit-R2 reconstructs operative intent via IC-CoT, then jointly optimizes IC-CoT generation and flow-matching image generation using unified multi-turn RL with prefix-valid advantage refinement.

In practice

• Implement IC-CoT to distill historical constraints.
• Employ trajectory filtering to prevent training noise.
• Utilize GA, IF, CC for multi-turn editing evaluation.

Topics

• Multi-turn Image Editing
• Reinforcement Learning
• In-Context Chain-of-Thought
• Multimodal Foundation Models
• MICE-Bench
• Global Awareness Metric

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing — Edit-R2 uses context-aware RL and intent reconstruction to enable robust multi-turn image editing.
• MT-EditFlow: Reinforcement Learning for Multi-Turn Image Editing with Flow Matching — MT-EditFlow optimizes multi-turn image editing using flow-matching reinforcement learning and a multi-reward formulation to mitigate error propagation.
• Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models — Aligning multi-turn LLM behavior with full-context teacher responses mitigates self-anchored drift.
• Multigrain-aware Semantic Prototype Scanning and Tri-Token Prompt Learning Embraced High-Order RWKV for Pan-Sharpening — A novel pan-sharpening method uses semantic-driven scanning and tri-token prompting with a high-order RWKV architecture.
• LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning — High-magnitude activations in query/key vectors are critical for long-context LLM reasoning and RL optimization.
• Multilingual OCR-Aware Fine-Tuning and Prompt-Guided Chain-of-Thought Reasoning for Multimodal Large Language Models — OCR-aware data curation and fine-tuning significantly improve MLLM robustness for multilingual text in degraded images.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.