Research map / Image generation / Editing and style transfer

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

NeurIPS 2026 · Image generation › Editing and style transfer · approach: Instruction-based diffusion editing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 Edit-R2, an RL post-training framework for multi-turn in-context image editing on BAGEL that combines in-context chain-of-thought intent reconstruction, unified Flow-GRPO/Text-GRPO optimization, and prefix-valid advantage refinement, plus the MICE-Bench benchmark.

Problem

Multi-turn in-context image editing, where a model must follow each new instruction while preserving session-level constraints from an interleaved image-text history; current models suffer long-context intent dilution and error propagation (state contamination) across turns.

Approach

Reconstruct the operative session intent via in-context chain-of-thought before each edit, then jointly optimize the reasoning tokens and flow-matching generation with a unified GRPO-style objective over multi-turn rollouts, truncating corrupted rollouts below a reward threshold and normalizing advantages only over active rollouts.

Key results

• Edit-R2 reaches 67.76 average at Turn 2 and 65.37 at Turn 3 on MICE-Bench, the best among open-source models, versus 58.91 and 58.61-level BAGEL variants.
• Edit-R2 improves BAGEL by about +18% IF and +18% GA at Turn 2 and +8% IF and +15% GA at Turn 3 on MICE-Bench.
• On GEdit-Bench-EN, Edit-R2 raises G_O from 6.52 (BAGEL) to 6.75.

Surprising findings

• An untrained IC-CoT produces excessive redundant reasoning that truncates output and degrades visual quality.
• Greedy-decoded VLM reward for IF fails to train at turn 1, degrading monotonically; stochastic self-ensemble (Avg@4) fixes this.

Limitations

• MICE-Bench covers only two in-context editing task categories (content memory and content understanding).
• IF and GA rely on VLM judges, introducing reward noise; GA human agreement is 81.5%.
• Edit-R2 gains on Turn 1 are modest relative to Qwen-Image-Edit's single-turn instruction following.

Experimental setup

Datasets: MICE-Bench training set (10K generated instances) EdiVal-Bench GEdit-Bench

Benchmarks: MICE-Bench EdiVal-Bench GEdit-Bench

Metrics: IF CC GA marginal task success rate Avg

Baselines: GPT-Image-1 Nano-Banana-2 Qwen-Image-Edit Flux.1 Kontext OmniGen VINCIE BAGEL BAGEL (w/ think)

Builds on: BAGEL Flow-GRPO Unified-GRPO EdiVal-Agent GroundingDINO

Closest papers by meaning

• The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.90)
• RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.89)
• Text-Vision Co-Instructed Image Editing (NeurIPS 2026 · similarity 0.88)
• VicEdit: Learning to Edit Videos from Visual In-Context Examples (NeurIPS 2026 · similarity 0.88)
• EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.87)
• Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.87)