FlowTool: Controlling Tool Parameter in Image Retouching via Flow Matching

The authors introduce FlowTool, a framework that frames tool-based image editing as a flow matching problem rather than relying on autoregressive MLLMs. By combining a vision-language model backbone with a Diffusion Transformer parameter generator, FlowTool directly models the distribution of continuous tool parameters from Gaussian noise given an input image and instruction. Trained via a two-stage supervised flow-matching curriculum and reward-based post-training, FlowTool outperforms specialized editing agents on reference-based benchmarks while drastically improving inference efficiency, reducing latency by at least 50times and requiring nearly 2times less memory than compared baselines.

FlowTool frames tool-based image editing as conditional generation over structured continuous editing parameters without autoregressive reasoning.

The framework significantly improves inference efficiency, reducing latency by at least 50times.

FlowTool requires nearly 2times less memory than the compared baselines.