FlowTool: Controlling Tool Parameter in Image Retouching via Flow Matching
FlowTool replaces step-by-step tool reasoning with direct generation of image-editing parameters.
The paper frames retouching as conditional flow matching over structured, continuous tool settings. Its system pairs a vision-language backbone with a Diffusion Transformer that turns noise into an editing plan. The authors report stronger reference-based results than specialized MLLM editing agents and proprietary MLLMs across four benchmarks, while staying competitive in reference-free evaluation. They also claim at least 50x lower latency and nearly 2x lower memory use than compared baselines. HF Daily Papers' note
The paper frames retouching as conditional flow matching over structured, continuous tool settings. Its system pairs a vision-language backbone with a Diffusion Transformer that turns noise into an editing plan. The authors report stronger reference-based results than specialized MLLM editing agents and proprietary MLLMs across four benchmarks, while staying competitive in reference-free evaluation. They also claim at least 50x lower latency and nearly 2x lower memory use than compared baselines. HF Daily Papers' note
score 4