Dotting the Eye: An Intent-Driven Image Retouching Agent for Visual Focus Enhancement

TL;DR

EyeControl is a multimodal large language model-driven agent that enhances visual focus in images using minimal user input like clicks or strokes. It links weak user intent to specific image regions and tonal adjustments via structured guidance and a diffusion-based retouching executor, improving attention alignment and retouching coherence. The system includes a novel operation-consistency constraint and a new benchmark for evaluation.

Why It Matters

It enables non-experts to achieve professional-level focus enhancement with minimal interaction, reducing reliance on manual skill. The approach advances intent-driven image editing by integrating semantic understanding with perceptual control, making retouching more intuitive and effective.

Key Findings

• EyeControl successfully translates sparse user inputs into precise visual focus adjustments using multimodal intent interpretation.
• The operation-consistency constraint improves coordination between global and local tonal changes, resulting in more natural retouching outcomes.
• ControlArt-Bench provides a high-quality dataset for evaluating focus enhancement, enabling more reliable benchmarking of intent alignment and visual quality.