Research map / Vision-language and multimodal / Multimodal LLMs

Meta-CoT: Enhancing Granularity and Generalization in Image Editing

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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 Meta-CoT, a two-level decomposition (triplet and meta-task) chain-of-thought for instruction-based image editing on Bagel, trained with SFT then Flow-GRPO using a VLM-based CoT-Editing Consistency reward.

Problem

Existing CoT-guided image editing in unified models either targets narrow understanding forms or fails to generalize across diverse editing tasks; the paper asks what CoT form and training strategy jointly improve understanding granularity and generalization.

Approach

Decomposes each editing operation into (task/meta-task, target, required understanding) triplets with a Task Summary, Task Thinking, and target-wise traversal CoT; reduces tasks to five meta-tasks (add, remove, replace, camera motion, position change) that compose into other tasks; and adds a VLM-judged CoT-editing consistency reward optimized with Flow-GRPO.

Key results

• Meta-CoT + RL improves the 21-task average to 6.415 vs 5.538 for Train Editing Only (about 15.8% gain as reported).
• On ImgEdit, Meta-CoT + RL reaches overall 3.83, above BAGEL(think) at 3.39 and BLIP3o-NEXT at 3.62.
• Training on only five meta-tasks achieves performance comparable to full-task training on the 21-task benchmark, per the ablation.

Surprising findings

• Bagel with thinking performs worse than without thinking on text editing (2.124 vs 5.170), and the Meta-CoT reasoning also hinders text editing.

Limitations

• Text editing degrades with extended textual reasoning, the only task where the method does not improve.
• Reliance on GPT-4.1 for evaluation and a VLM reward whose alignment was calibrated against four annotators on 500 samples.
• Built on a single base model (Bagel), so generality to other unified models is not shown.

Experimental setup

Datasets: LLaVA-OneVision Mammoth-VL GEdit-Bench RiseBench ComplexEdit

Benchmarks: 21-task editing benchmark ImgEdit

Metrics: VIEScore overall instruction following consistency naturalness artifacts ImgEdit overall score

Baselines: Bagel (w/o think) Bagel (w think) Train Editing Only FLUX.1 Kontext [Pro] GPT Image 1 Qwen-Image OmniGen2 BLIP3o-NEXT

Builds on: Bagel Flow-GRPO GoT Gedit-Bench ImgEdit

Closest papers by meaning

• Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.89)
• ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.89)
• From Scale to Speed: Adaptive Test-Time Scaling for Image Editing (CVPR 2026 · similarity 0.89)
• EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.88)
• KRIS-Bench: Benchmarking Next-Level Intelligent Image Editing Models (NeurIPS 2025 · similarity 0.88)
• UnicEdit-10M: A Dataset and Benchmark Breaking the Scale-Quality Barrier via Unified Verification for Reasoning-Enriched Edits (CVPR 2026 · similarity 0.88)