Research map / Vision-language and multimodal / Multimodal LLMs

RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 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 RetouchIQ, an MLLM agent that turns natural-language retouching instructions into parameterized Lightroom edits, trained with SFT and RL using a generalist reward model (GRM) refined by policy-guided reward training (PGRT).

Problem

Instruction-based image retouching with professional editing tools needs reward signals, but rule-based pixel/reference-similarity rewards are unreliable because multiple edits can be equally valid and subjective.

Approach

Replace fixed reference-based rewards with an RL-fine-tuned MLLM reward model that generates case-specific evaluation metrics and then scores edits, and align its training data to the policy's own edit distribution via PGRT (policy-generated weak edits instead of perturbed ones), with alternating policy/reward training.

Key results

• On RetouchEval quality improving, RetouchIQ-GRM reaches SC 7.57 and PQ 7.48 versus JarvisArt's 7.22 and 6.59.
• On MIT-Adobe5K, RetouchIQ-GRM achieves SSIM 0.86, LPIPS 0.16, PSNR 23.14, versus MonetGPT's 0.82, 0.17, 23.10.
• PGRT is reported to raise reward-model accuracy on policy-generated weak edits and yield the best policy performance among reward-model variants.

Surprising findings

• Verifiable rewards and perturbation-based reward training perform sub-optimally, since perturbed weak edits (mostly single exposure/temperature changes) differ in distribution from policy outputs.

Limitations

• Evaluation relies on GLM-4.5V as judge for perceptual and semantic scores, and RetouchEval is a 300-sample self-curated benchmark.
• Compares against closed-source general MLLMs and a single open diffusion model (FLUX-Pro) rather than other retouching systems.

Experimental setup

Datasets: user editing trajectories from real users (before/after pairs) MIT-Adobe5K

Benchmarks: RetouchEval MIT-Adobe5K

Metrics: L1 L2 SC (semantic consistency) PQ (perceptual quality) O (overall) SSIM LPIPS PSNR

Baselines: GPT-5 Gemini-2.5 MonetGPT JarvisArt Flux-Pro

Builds on: JarvisArt MonetGPT Qwen2.5-VL

Closest papers by meaning

• Guiding Instruction-based Image Editing via Multimodal Large Language Models (ICLR 2024 · similarity 0.91)
• JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.90)
• Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.89)
• The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.89)
• ReasonEdit: Towards Reasoning-Enhanced Image Editing Models (CVPR 2026 · similarity 0.89)
• SmartEdit: Exploring Complex Instruction-based Image Editing with Multimodal Large Language Models (CVPR 2024 · similarity 0.89)