FlowTool: Controlling Tool Parameter in Image Retouching via Flow Matching
Thanh-Long V. Le, Steven Walton, Seunghyun Yoon, Branislav Kveton, Trung Bui, Eunho Yang, Viet Lai
Adobe Research · KAIST
Published September 28, 2026 · on the feed September 29, 2026 · ▲ 31 · arXiv 2609.35673
The abstract
Tool-based image editing (image retouching) is commonly formulated with autoregressive multimodal large language models (MLLMs) that sequentially generate reasoning, tool selections, and parameter values. In this work, we present a novel approach to tool-based image editing by framing the task as a flow matching problem. We introduce FlowTool, a framework that directly models the distribution of high-quality tool parameters conditioned on the input image and user instruction using conditional rectified flow. FlowTool combines a vision-language model backbone for multimodal understanding with a Diffusion Transformer parameter generator that transforms Gaussian noise into an editing plan. We train FlowTool with a two-stage supervised flow-matching curriculum, followed by reward-based post-training. Across MMArt-Bench, FlowTool-Eval, ArtEdit-Bench, and MIT-Adobe5K, FlowTool achieves significantly stronger reference-based performance than specialized MLLM editing agents and proprietary MLLMs, while remaining competitive with proprietary models under reference-free evaluation. Moreover, FlowTool significantly improves inference efficiency, reducing latency by at least 50times while requiring nearly 2times less memory than the compared baselines. These results demonstrate that tool-based image editing can be effectively modeled as conditional generation over structured continuous editing parameters, without autoregressive reasoning.
The Brief
2:30A short podcast summary of the paper. Written and voiced by AI.
Transcript
FlowTool. Controlling Tool Parameter in Image Retouching via Flow Matching. By Thanh-Long V. Le and six others, from Adobe Research and KAIST. Published on the twenty-eighth of September. No code repository is linked to it.
Image retouching with tools, where an edit is expressed as a choice of tools and the values of their parameters, is usually handled by autoregressive multimodal large language models. They generate in sequence: first reasoning, then the tool selections, then the parameter values, token by token. The output, though, is ultimately a set of continuous settings, and producing them this way means running the whole chain of reasoning each time.
FlowTool treats the task instead as a flow matching problem. It directly models the distribution of high-quality tool parameters, conditioned on the input image and the user's instruction, using conditional rectified flow. An editing plan is not written out step by step. It is generated as a whole.
Two components divide the work. A vision-language model backbone handles multimodal understanding of the image and the request. A Diffusion Transformer parameter generator then turns Gaussian noise into an editing plan. Training runs in two supervised flow-matching stages, arranged as a curriculum, and is followed by reward-based post-training. The division is clean. Understanding what the user wants and what the picture looks like belongs to the vision-language model. Turning that understanding into concrete settings belongs to the flow-based generator, which starts from noise and moves it, under the conditioning, toward a plan that reflects the high-quality parameters it was trained on.
Evaluation covers MMArt-Bench, FlowTool-Eval, ArtEdit-Bench and MIT-Adobe5K. Under reference-based evaluation, FlowTool is significantly stronger than specialized multimodal editing agents and than proprietary models. Under reference-free evaluation it remains competitive with the proprietary models. The efficiency numbers are larger still. Latency drops by at least fifty times, and memory use is nearly halved compared with the baselines. The quality picture depends on the kind of evaluation: clearly ahead when there is a reference edit to compare against, level with proprietary models when there is not.
The authors read these results as evidence that tool-based editing can be modeled as conditional generation over structured, continuous editing parameters, without any autoregressive reasoning. If a retouching plan is really a point in a space of settings, a model that samples that point directly can be both better and far cheaper than one that talks its way there.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A slider is a number, not a word. That mismatch is where this quest begins.
You build an assistant for a professional photo editor. A user types 'make it more cinematic', and the assistant must pick editing sliders and set each one to a value.
The road ahead
- 1 · The problem ~7 min Photo-editing assistants that write slider values as text tokens get the numbers imprecise, carry early mistakes forward, and run slowly.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Treat the edit as one set of continuous slider values, let a vision-language model understand the request, and let a Diffusion Transformer generate the values from noise.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Encode photo and instruction with a vision-language model, turn noise into tool parameters with a Diffusion Transformer, then apply them with the real editing tools.
- 6 · The experiment ~6 min You’ll guess: By generating tool parameters with flow matching instead of autoregressive reasoning, how much did FlowTool cut latency compared with the baselines?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.