Recent advances in multimodal large language models (MLLMs) have shown great potential for extending vision-language reasoning to professional tool-based image editing, enabling intuitive and creative editing. A promising direction is to use reinforcement learning (RL) to enable MLLMs to reason about and execute optimal tool-use plans within professional image-editing software. However, training remains challenging due to the lack of reliable, verifiable reward signals that can reflect the inherently subjective nature of creative editing. In this work, we introduce RetouchIQ, a framework that performs instruction-based executable image editing through MLLM agents guided by a generalist reward model. RetouchIQ interprets user-specified editing intentions and generates corresponding, executable image adjustments, bridging high-level aesthetic goals with precise parameter control. To move beyond conventional, rule-based rewards that compute similarity against a fixed reference image using handcrafted metrics, we propose a generalist reward model, an RL fine-tuned MLLM that evaluates edited results through a set of generated metrics on a case-by-case basis. Then, the reward model provides scalar feedback through multimodal reasoning, enabling reinforcement learning with high-quality, instruction-consistent gradients. We curate an extended dataset with 190k instruction-reasoning pairs and establish a new benchmark for instruction-based image editing. Experiments show that RetouchIQ substantially improves both semantic consistency and perceptual quality over previous MLLM-based and diffusion-based editing systems. Our findings demonstrate the potential of generalist reward-driven MLLM agents as flexible, explainable, and executable assistants for professional image editing.
Figures & tables
Figure 1: Overview of RetouchIQ . Left: We annotate the user instruction and reasoning for training data. Generated data are filtered to ensure quality. Middle: The supervised fine-tuning stage. Based on the user instruction ( e.g., pop more), the policy model needs to reason the correct parameters ( e.g., exposure ) and change them accordingly. Right: The reinforcement learning stage. We leverage a generalist reward model to propose metrics and provide scalar reward guidance for policy model. Details of the reward model is introduced in Sec 4 .
Figure 2: Problematic rewards in image retouching tasks. Given a before–after image pair ( left ), ❶ verifiable rewards ( middle ) rely on metrics between the edited image and ground truth, such as pixel differences. However, since multiple valid edits can satisfy user intent, these rewards become imprecise. ❷ The reward model ’s precision strongly depends on its training data distribution ( right ). When trained to distinguish good user edits from randomly perturbed images, it may later struggle to assess results from the policy model that produces combined, complex edits.
Figure 3: Overview of generalist reward model. Left : Given a before-edited image and a user-edited after image After-u (“u” for “user”), we perturb the editing process to obtain a suboptimal image, After-w (“w” for “weak”). Middle : In the SFT stage,the reward model first learns to generate metrics and then produces scalar rewards for the two input images. Right : In the RL stage, we introduce policy-guided reward training (PGRT), where the suboptimal image After-w is provided by the policy model rather than generated through perturbation. The goal of the RL stage is to assign a higher scalar reward to the user-edited (strong) image After-u .
Quality Improving
Style Changing
Local Retouching
L1 ↓
L2 ↓
SC ↑
PQ ↑
O ↑
L1 ↓
L2 ↓
SC ↑
PQ ↑
O ↑
L1 ↓
L2 ↓
SC ↑
PQ ↑
O ↑
Flux-Pro [ 21 ]
64.97
98.14
6.12
6.09
6.10
68.20
70.01
6.48
6.36
6.42
67.93
82.86
6.19
6.26
6.22
GPT-5 [ 27 ]
36.21
47.53
6.74
6.51
6.62
39.17
50.36
6.89
6.76
6.82
37.20
47.56
6.07
6.89
6.47
Gemini-2.5 [ 6 ]
39.58
51.60
6.56
6.72
6.64
41.54
53.01
5.82
6.28
6.05
36.11
46.72
6.01
6.63
6.31
MonetGPT [ 10 ]
32.05
44.98
6.52
7.05
6.78
41.03
52.75
5.87
6.03
5.95
30.08
44.98
6.27
6.70
6.48
JarvisArt [ 22 ]
32.17
46.23
7.22
6.59
6.90
33.90
48.23
7.39
6.88
7.13
38.19
49.75
6.45
6.39
6.42
Table 1: Quantitative comparisons on the RetouchEval benchmark. We consider three categories of baselines: general-purpose MLLMs (GPT-5, Gemini-2.5); MLLM agents (MonetGPT, JarvisArt); and diffusion-based methods (Flux-Pro). Meanwhile, we report the performance of RetouchIQ under both the SFT (RetouchIQ-SFT) and RL (RetouchIQ-RL) stages. We also include results from a variant trained with rule-based reward (RetouchIQ-Rule). The best results are shown in bold , and the second-best results are underlined .
Figure 4: Comparison of reward model and policy model performance under different reward model configurations. The lines show the accuracies of the reward model, while the bars indicate the scores of the corresponding policy model.
Figure 5: Qualitative results across diverse image retouching scenarios, including quality enhancement (top), style transformation (middle), and local retouching (bottom). For each example, the input image and the corresponding user-edited result are shown on the left.
SSIM
LPIPS
PSNR
GPT-5 [ 27 ]
0.72
0.26
20.82
MonetGPT [ 10 ]
0.82
0.17
23.10
JarvisArt [ 22 ]
0.76
0.23
21.03
RetouchIQ-SFT
0.84
0.20
22.37
RetouchIQ-GRM
0.86
0.16
23.14
Table 2: Quantitative results on the MIT-Adobe5K benchmark.
Current image editing software often hinges on fixed filters or expert tuning, leaving a gap between amateur users' intent and outcomes. Creations by generative models may contain artifacts, implausible details, or stylistic drift away from photorealism and offer little insight into why an edit was made. We propose IEA, a conversational Image Editing Agent that learns to operate parameterized tools in an explicit, interpretable action space. IEA is trained via a three-stage multitask pipeline: (1) SFT on distilled expert edits, (2) GRPO with rewards for likeness improvement, tool usefulness, and intent summarization, and (3) large-scale synthetic fine-tuning to jointly master image editing, refinement, and user intent summarization. By manipulating 16 editing tools step by step, IEA produces transparent edit traces that can be inspected and debugged. In quantitative experiments, it attains a lower pixel distance on the edit task and a higher ROUGE-L on the summary task than strong baselines. In user studies, it ranks best among tool-calling methods for instruction following while surpassing generative methods in overall perceptual quality. Our results validate interpretable, tool-centric VLMs as a reliable path to human instruction-guided image retouching.
Zichen Zhu, Yuheng Sun, Mingxuan Zhu +11
X-LANCE Lab, School of Computer Science, Shanghai Jiao Tong University · Huawei Technologies Ltd. · Shanghai Innovation Institution +1
Existing instruction-based image editing models perform well with simple, single-step instructions but degrade in realistic scenarios that involve multiple, lengthy, and interdependent directives. A main cause is the scarcity of training data with complex multi-instruction annotations. However, it is costly to collect such data and retrain these models. To address this challenge, we propose MSRAMIE, a training-free agent framework built on Multimodal Large Language Model (MLLM). MSRAMIE takes existing editing models as plug-in components and handle multi-instruction tasks via structured multimodal reasoning. It orchestrates iterative interactions between an MLLM-based Instructor and an image editing Actor, introducing a novel reasoning topology that comprises the proposed Tree-of-States and Graph-of-References. During inference, complex instructions are decomposed into multiple editing steps which enable state transitions, cross-step information aggregation, and original input recall, which enables systematic exploration of the image editing space and flexible progressive output refinement. The visualizable inference topology further provides interpretable and controllable decision pathways. Experiments show that as the instruction complexity increases, MSRAMIE can improve instruction following over 15% and increases the probability of finishing all modifications in a single run over 100%, while preserving perceptual quality and maintaining visual consistency.
Zhaoyuan Qiu, Ken Chen, Xiangwei Wang +3
University Of Melbourne, Grattan Street, Parkville, Victoria 3010, Australia
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 50× while requiring nearly 2× 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.
Thanh-Long V. Le, Steven Walton, Seunghyun Yoon +4