Reproduce “RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qiucheng Wu, Jing Shi, Simon Jenni, Kushal Kafle, Tianyu Wang, Shiyu Chang, Handong Zhao
RETOUCHIQ presents an MLLM agent designed for instruction-based image retouching, translating natural language into precise, executable operations within professional software. The system addresses the inherent subjectivity of creative editing tasks through a Generalist Reward Model and Policy-Guided Reward Training, producing professional-grade, instruction-aligned image edits.
Key insights
- MLLMs can be effectively leveraged as executable agents that translate high-level natural language instructions into precise, controllable parameter adjustments for professional-grade creative software.
- Subjective creative tasks benefit significantly from a dynamically generated, context-aware reward model (GRM) that evaluates edits based on self-generated metrics, providing reliable feedback for reinforcement learning.
- Aligning the reward model's training data distribution with the policy's evolving outputs through Policy-Guided Reward Training (PGRT) is critical for stable and effective RL, resolving issues of distribution shift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.