Organizations: Shanghai Jiao Tong University · Shanghai AI Laboratory · CUHK MMLab · Fudan University · University of Science and Technology of China · Xiamen University
Instruction-based image editing has emerged as a key capability for unified multimodal models (UMMs), yet constructing large-scale, diverse, and high-quality editing datasets without costly proprietary APIs remains challenging. Previous image editing datasets either rely on closed-source models for annotation, which prevents cost-effective scaling, or employ fixed synthetic editing pipelines, which suffer from limited quality and generalizability. To address these challenges, we propose ScaleEditor, a fully open-source hierarchical multi-agent framework for end-to-end construction of large-scale, high-quality image editing datasets. Our pipeline consists of three key components: source image expansion with world-knowledge infusion, adaptive multi-agent editing instruction-image synthesis, and a task-aware data quality verification mechanism. Using ScaleEditor, we curate ScaleEdit-12M, the largest open-source image editing dataset to date, spanning 23 task families across diverse real and synthetic domains. Fine-tuning UniWorld-V1 and Bagel on ScaleEdit yields consistent gains, improving performance by up to 10.4% on ImgEdit and 35.1% on GEdit for general editing benchmarks and by up to 150.0% on RISE and 26.5% on KRIS-Bench for knowledge-infused benchmarks. These results demonstrate that open-source, agentic pipelines can approach commercial-grade data quality while retaining cost-effectiveness and scalability. Both the framework and dataset will be open-sourced.
Figures & tables
Figure 1: Overview of ScaleEdit. (a) Examples of instruction-based image editing in ScaleEdit, covering diverse local and global editing types. The central wheel illustrates the taxonomy of editing categories supported by ScaleEdit. (b) Size comparison between ScaleEdit and existing instruction-based image editing datasets constructed from open-source and commercial models. (c) Performance of UniWorld-V1 [ 21 ] before and after fine-tuning on ScaleEdit across multiple benchmarks (GEdit-EN-full [ 24 ] , ImgEdit-Full [ 43 ] , KRIS-Bench [ 39 ] , RISEBench [ 51 ] ), where ScaleEdit consistently brings significant improvements.
Figure 2: Overview of ScaleEditor. The framework consists of three main stages: (1) Source Image Expansion, which retrieves and synthesizes diverse images while filtering and deduplicating noisy samples; (2) Adaptive Multi-Agent Synthesis, where a task router dispatches each image to appropriate editing workflows, with specialized agents to synthesize corresponding editing instruction and images; and (3) Task-Aware Quality Verification, which assesses instruction alignment, editing consistency, and generation quality. These high-quality instruction-image-edit triplets form our ScaleEdit.
Figure 3: Diversity and quality analysis of ScaleEdit. (a) Distribution of editing subcategories demonstrates the broad coverage of editing behaviors in ScaleEdit. (b) Word cloud of editing instructions highlights the linguistic richness in edit semantics. (c) Multi-dimensional filtering score distribution across three evaluation metrics. (d) Comparison of filtering scores across representative datasets (ShareGPT-4o [ 3 ] , UltraEdit [ 50 ] , AnyEdit [ 45 ] ), where each sample's final score is defined as the minimum of its three evaluation metrics.
Model
Background
Color
Material
Motion
Portrait
Style
Add
Remove
Replace
Text
Tone
Avg
GPT-4o [ 27 ]
6.96
6.85
7.10
5.41
6.74
7.44
7.51
8.73
8.55
8.45
8.69
7.49
OmniGen [ 40 ]
5.23
5.93
5.44
3.12
3.17
4.88
6.33
6.35
5.34
4.31
4.96
5.01
Step1X-Edit [ 24 ]
7.03
6.26
6.46
3.66
5.23
7.24
7.17
6.42
7.39
7.40
6.62
6.44
Finetuning on UniWorld-V1
▼ Baseline
UniWorld-V1 [ 21 ]
4.92
6.37
4.79
1.85
4.03
5.64
7.23
6.17
5.70
1.15
5.54
4.85
▼ w/ Commercial Datasets
Table 1: Comparison of fine-tuning results on GEdit-EN-Full [ 24 ] . For each baseline, the best score is shown in bold .
Model
Add
Adjust
Extract
Replace
Remove
Background
Style
Hybrid
Action
Avg
GPT-4o [ 27 ]
4.61
4.33
2.90
4.35
3.66
4.57
4.93
3.96
4.89
4.20
OmniGen [ 40 ]
3.47
3.04
1.71
2.94
2.43
3.21
4.19
2.24
3.38
2.96
Step1X-Edit [ 24 ]
3.88
3.14
1.76
3.40
2.41
3.16
4.63
2.64
2.52
3.06
Finetuning on UniWorld-V1
▼ Baseline
UniWorld-V1 [ 21 ]
3.82
3.64
2.27
3.47
3.24
2.99
4.21
2.96
2.74
3.26
▼ w/ Commercial Datasets
Table 2: Comparison of fine-tuning results on ImgEdit-Bench [ 43 ] . For each baseline, the best score is shown in bold .
Model
RISEBench [ 51 ]
KRIS Bench [ 39 ]
Reasoning
ApprConsistency
VisualPlausibility
Overall
Factual
Conceptual
Procedural
Overall
GPT-4o [ 27 ]
62.80
80.20
94.90
28.90
79.80
81.37
78.32
80.09
OmniGen [ 40 ]
22.00
32.60
55.30
0.80
33.11
28.02
23.89
28.85
Step1X-Edit [ 24 ]
25.10
41.50
73.50
1.90
45.52
48.01
31.82
43.29
Finetuning on UniWorld-V1
▼ Baseline
UniWorld-V1 [ 21 ]
18.33
65.79
86.63
2.22
47.71
44.80
47.92
50.27
Table 3: Comparison of fine-tuning results on RISEBench [ 51 ] and KRIS Bench [ 39 ] . For each baseline, the best score is shown in bold .
Training Data
UniWorld-V1 [ 21 ]
Janus-Pro [ 5 ]
ImgEdit [ 43 ]
GEdit [ 24 ]
ImgEdit [ 43 ]
GEdit [ 24 ]
▼ Baseline (No Fine-tuning)
Original
3.26
4.85
–
–
▼ w/ Commercial Datasets
OpenGPT-4o-Image [ 6 ]
3.37
6.06
3.08
4.52
ShareGPT-4o-Image [ 3 ]
3.49
6.07
2.96
4.47
Table 4: Ablation study of different datasets with equal scale. All experiments are conducted under the exact same settings and data size for a fair comparison. The best score is shown in bold , and the second best is underlined .
Model
Task Router
Data Filtering
ImgEdit [ 43 ]
GEdit [ 24 ]
Janus-Pro [ 5 ]
✓
✗
3.09
4.82
✗
✓
3.14
4.85
✓
✓
3.17
4.92
UniWorld-V1 [ 21 ]
✓
✗
3.36
6.06
✗
✓
3.43
6.07
✓
✓
3.50
6.15
Table 5: Ablation study on Data Filtering and Task Router. Our routing and filtering strategy improves performance on general editing tasks across different models.
Model
Training Data
RISEBench [ 51 ]
Reasoning
Consistency
VisPlaus.
Overall
UniWorld-V1 [ 21 ]
w/o rewrite
22.75
57.35
90.26
3.05
w/ rewrite
24.38
64.24
90.64
4.17
Table 6: Ablation study on instruction rewriting. VisPlaus. denotes the Visual Plausibility sub-dimension.
Models
Accuracy
MAE
Instruction Following
Editing Consistency
Generation Quality
Instruction Following
Editing Consistency
Generation Quality
InternVL3-72B [ 52 ]
0.81
0.61
0.85
0.22
0.34
0.22
Qwen2.5-VL-72B [ 1 ]
0.82
0.63
0.89
0.17
0.28
0.17
Table 7: Alignment of open-source MLLM judges with GPT-4o. We compare the scoring consistency of different models against GPT-4o across three key quality dimensions. Higher Accuracy and lower MAE indicate better alignment with GPT-4o. Qwen2.5-VL-72B consistently outperforms InternVL3-72B across all metrics.
Models
Accuracy
MAE
Instruction Following
Editing Consistency
Generation Quality
Instruction Following
Editing Consistency
Generation Quality
Qwen2.5-VL-72B [ 1 ]
0.78
0.67
0.78
0.24
0.31
0.26
GPT-4o [ 27 ]
0.86
0.75
0.82
0.16
0.22
0.20
Table 8: Alignment of MLLM judges with human preference. Higher Accuracy and lower MAE indicate better alignment with human preference. Qwen2.5-VL-72B shows highly competitive alignment with human evaluators, demonstrating its reliability at scale compared to GPT-4o.
Models
GEdit [ 24 ]
ImgEdit [ 43 ]
Baseline
Finetuned
Baseline
Finetuned
InstructPix2Pix [ 2 ]
3.68
3.78
1.88
2.06
OmniGen [ 40 ]
5.06
5.49
2.96
3.15
Step1X-Edit [ 24 ]
6.70
7.14
3.06
3.21
Table 9: Finetuned results across different models. All experiments are conducted using the same settings as the equal-scale comparison.
Figure 4: Qualitative comparison of UniWorld-V1 before and after fine-tuning on ScaleEdit.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Overview of the Source Image Expansion stage in ScaleEditor. Retrieval-based expansion performs image-based and text-based retrieval to collect semantically related images, while synthesis-based expansion produces high-fidelity variants, guided by detailed image captions.
Category
Editing Task
Definition
Global-level Editing
Style Transfer
Converts the overall artistic appearance of an image into a target style while preserving its structural and semantic content.
Tone Adjustment
Adjusts global tonal parameters such as brightness, contrast, saturation, or color temperature.
Viewpoint Transformation
Modifies the camera viewpoint or perspective geometry to present the scene from a new spatial angle.
Background Replacement
Replaces the entire background of an image while maintaining the integrity of the foreground objects.
Object-level Editing
Object Addition
Inserts new object instances into the scene while maintaining coherent spatial relationships and lighting conditions.
Object Removal
Removes designated objects and reconstructs the occluded background to preserve scene realism and continuity.
Appendix
Table 10: Detailed definition of editing tasks .
Dataset
Instruction Alignment
Edit Consistency
Generation Quality
1
2
3
1
2
3
1
2
3
ScaleEdit
0.0%
0.0%
100.0%
0.0%
14.2%
85.8%
0.0%
23.5%
76.5%
ScaleEdit (pre-filtered)
7.5%
9.7%
82.8%
1.6%
14.0%
84.4%
1.2%
23.2%
75.6%
ShareGPT-4o-Image [ 3 ]
3.9%
8.3%
87.8%
0.4%
9.7%
89.9%
0.5%
12.1%
87.4%
OmniEdit [ 37 ]
5.8%
13.0%
81.2%
0.4%
24.8%
74.8%
0.5%
25.3%
74.2%
ImgEdit [ 43 ]
19.9%
25.6%
54.5%
3.7%
38.4%
58.0%
2.4%
49.2%
48.4%
Appendix
Table 11: Score distribution of each dimension across different datasets.
Dataset
(3,3,3)
(3,3,2)
(2,2,2)
(1,2,2)
(3,2,2)
(2,3,2)
(2,3,3)
(3,2,3)
(2,2,3)
(1,1,1)
ScaleEdit
85.3%
10.4%
0.0%
0.0%
2.7%
0.0%
0.0%
1.7%
0.0%
0.0%
ScaleEdit (pre-filtered)
70.6%
8.6%
4.5%
4.2%
2.2%
2.4%
1.8%
1.4%
1.1%
0.9%
ShareGPT-4o-Image [ 3 ]
81.3%
3.7%
3.3%
2.1%
1.2%
1.4%
2.5%
1.6%
1.1%
0.3%
OmniEdit [ 37 ]
63.5%
7.3%
7.6%
4.1%
4.4%
1.4%
1.9%
6.0%
2.1%
0.3%
ImgEdit [ 43 ]
37.5%
9.0%
15.3%
11.5%
6.1%
4.2%
3.8%
1.9%
2.3%
2.0%
UltraEdit [ 50 ]
44.7%
19.2%
9.1%
7.1%
11.4%
1.3%
1.0%
1.9%
0.8%
1.6%
Appendix
Table 12: The distribution of joint score combinations across datasets. Each tuple denotes (F, C, Q) .
Edit Subtask
Top-1 (%)
Top-5 (%)
Top-10 (%)
Top-20 (%)
Vocab Size
▼ Open-Ended Semantic Editing
Background Replacement
9.81
22.21
29.58
38.81
2456
Object Addition
1.48
5.53
8.86
12.50
6222
Object Removal
10.19
21.41
25.44
30.03
3987
Object Replacement
5.17
10.19
12.60
16.01
5590
Action Editing
2.54
10.12
14.61
21.09
3764
Appendix
Table 13: Target Entity Distribution. The Top- K columns indicate the cumulative percentage of the K most frequent entities. Vocab Size represents the total number of unique core entities extracted for that task. The data demonstrates a healthy long-tail distribution for open-ended tasks and extreme diversity for context-driven tasks.
Figure 6: T-SNE projection of semantic diversity across editing tasks. The absence of highly isolated or overly dense clusters demonstrates our dataset’s semantic richness and avoids repetitive or formulaic instruction generation.
Ratio
Percentage
Ratio
Percentage
Ratio
Percentage
Ratio
Percentage
0.56
0.57%
0.65
0.39%
0.66
1.36%
0.67
10.47%
0.70
0.42%
0.71
0.61%
0.80
0.77%
1.00
8.81%
1.25
0.41%
1.33
8.92%
1.34
0.44%
1.36
0.40%
1.39
0.59%
1.40
0.38%
1.49
1.13%
1.50
34.67%
1.51
3.00%
1.60
0.40%
1.78
4.56%
–
–
Appendix
Table 14: Aspect Ratio Distribution.
Parameters
UniWorld-V1 [ 21 ]
Bagel [ 9 ]
batch size
128
256
iterations
89640
44820
learning rate
10−5
2×10−5
warmup rate
0.05
0.05
optimizer
AdamW
AdamW
LR scheduler
Const
Const
Appendix
Table 15: Training Parameters for UniWorld-V1 [ 21 ] and Bagel [ 9 ] . LR scheduler denotes learning rate scheduler.
Parameters
UniWorld-V1 [ 21 ]
Janus-Pro [ 5 ]
batch size
128
32
iterations
7,812
31,250
gradient accumulation
1
4
learning rate
10−5
2×10−6
warmup rate
0.05
0.05
optimizer
AdamW
AdamW
Appendix
Table 16: Training Parameters for Ablation Studies. LR scheduler denotes learning rate scheduler.
Model
Training Dataset
RISEBench [ 51 ]
KRIS Bench [ 39 ]
Reasoning
ApprCons.
VisPlausi.
Overall
Factual
Conceptual
Procedural
Overall
UniWorld-V1 [ 21 ]
w/o rewrite
22.75
57.35
90.26
3.05
54.86
59.65
34.73
52.19
w/ rewrite
24.38
64.24
90.64
4.17
56.83
63.15
35.99
54.88
Appendix
Table 17: Ablation on instruction rewriting. ApprCons. denotes Appearance Consistency sub-dimension, while VisPlausi. represents Visual Plausibility sub-dimension.
Dataset
Background
Color
Material
Motion
Portrait
Style
Add
Remove
Replace
Text
Tone
Avg
UniWorld-V1 [ 21 ]
Pre-trained
4.92
6.37
4.79
1.85
4.03
5.64
7.23
6.17
5.70
1.15
5.54
4.85
▼ w/ Commercial Datasets
OpenGPT-4o-Image [ 6 ]
6.88
6.53
6.24
6.48
6.07
7.11
7.01
6.19
5.81
2.23
6.09
6.06
ShareGPT-4o-Image [ 3 ]
6.42
7.60
6.43
6.16
6.26
6.62
7.26
5.85
5.31
2.26
6.60
6.07
Nano-consistent [ 42 ]
6.76
6.10
4.23
5.55
6.31
5.41
6.44
1.44
5.24
1.36
6.39
5.02
Pico-Banana [ 30 ]
6.58
7.66
5.25
6.93
6.42
5.36
6.57
6.10
6.01
2.30
6.77
5.99
Appendix
Table 18: Ablation study of different datasets with equal scale on GEdit-EN-Full [ 24 ] The best score is shown in bold , and the second best is underlined , for UniWorld-V1 and Janus-Pro, respectively.
Dataset
Add
Adjust
Extract
Replace
Remove
Background
Style
Hybrid
Action
Avg
UniWorld-V1 [ 21 ]
Pre-trained
3.82
3.64
2.27
3.47
3.24
2.99
4.21
2.96
2.74
3.26
▼ w/ Commercial Datasets
OpenGPT-4o-Image [ 6 ]
4.18
3.96
1.99
3.44
2.62
3.67
4.65
2.74
3.07
3.37
ShareGPT-4o-Image [ 3 ]
4.11
3.67
2.04
3.53
2.90
3.83
4.81
2.94
3.62
3.49
Nano-consistent [ 42 ]
3.88
3.52
1.91
3.23
2.21
3.11
4.07
2.83
4.10
3.21
Pico-Banana [ 30 ]
4.08
3.79
1.86
3.71
3.56
3.56
4.11
2.85
3.26
3.42
Appendix
Table 19: Ablation study of different datasets with equal scale on ImgEdit-Bench. [ 43 ] The best score is shown in bold , and the second best is underlined , for UniWorld-V1 and Janus-Pro, respectively.
Figure 7: Qualitative comparison of Bagel [ 9 ] before and after fine-tuning on ScaleEdit.
Figure 8: User interface used in the user study. The left panel displays the original image and the corresponding editing instruction, while the right panel shows four candidate edited results from different models, labeled as options A, B, C, and D. Annotators are asked to select the option that best satisfies the editing instruction.
Training Data
Top-1 preference
GPT-Image-Edit [ 35 ]
26.8%
OpenGPT-4o-Image [ 6 ]
28.1%
ImgEdit [ 43 ]
15.6%
ScaleEdit
29.5%
Appendix
Table 20: Human preference results in the user study. “Top-1 preference” denotes the percentage of tasks for which each model is selected as the best edited result.
Figure 9: Examples corresponding to each score of different metrics.
Figure 10: Examples across different editing tasks.
Existing image editing frameworks predominantly follow the training paradigm of text-to-image diffusion models. However, extending this paradigm to image editing highlights two inherent discrepancies, specifically, the insufficient attention to edit concept granularity and the training inefficiency caused by sparse supervision signals. To address these issues, we establish a comprehensive hierarchical taxonomy featuring over 1,000 fine-grained edit concepts and build ConceptEdit-12M, a massive dataset of 12 million high-quality editing pairs via an improved synthesis framework. This library-driven approach effectively rectifies the distribution collapse of generated data while ensuring high data fidelity. Furthermore, we propose a dense supervision training strategy that synthesizes multiple non-interfering concepts into single image pairs. By providing richer learning signals, this strategy significantly enhances both training efficiency and overall model performance. Training results validate our strategy, significantly outperforming prior works. Finally, we present ConceptEdit-Bench, a granular evaluation suite designed to diagnose model capabilities across a vast array of real-world scenarios.
With the rapid advancement of commercial multi-modal models, image editing has garnered significant attention due to its widespread applicability in daily life. Despite impressive progress, existing image editing systems, particularly closed-source or proprietary models, often struggle with complex, indirect, or multi-step user instructions. These limitations hinder their ability to perform nuanced, context-aware edits that align with human intent. In this work, we propose ImageEdit-R1, a multi-agent framework for intelligent image editing that leverages reinforcement learning to coordinate high-level decision-making across a set of specialized, pretrained vision-language and generative agents. Each agent is responsible for distinct capabilities--such as understanding user intent, identifying regions of interest, selecting appropriate editing actions, and synthesizing visual content--while reinforcement learning governs their collaboration to ensure coherent and goal-directed behavior. Unlike existing approaches that rely on monolithic models or hand-crafted pipelines, our method treats image editing as a sequential decision-making problem, enabling dynamic and context-aware editing strategies. Experimental results demonstrate that ImageEdit-R1 consistently outperforms both individual closed-source diffusion models and alternative multi-agent framework baselines across multiple image editing datasets.
Image generation and editing models have advanced rapidly, yet remain unreliable when prompts require external world knowledge. Bounded and long-tail parametric knowledge prevents direct or reason-then-generate approaches from recovering the required facts and visual appearances. Existing agentic generation and editing methods mitigate this limitation with retrieval tools, yet remain constrained by insufficient visual verification, overloaded policy models, and weak integration of retrieved textual and visual evidence. To address these limitations, we present WeAgent-MMGenEdit, a full-stack recipe including a multimodal harness, a scalable data construction pipeline, a comprehensive benchmark, and post-training methods for the agent policy and image backend. We first introduce WeAgent-Harness, a multimodal runtime with persistent evidence management and dedicated verification and integration tools that organize retrieved multimodal evidence into a dense carrier. Upon this, we develop a scalable pipeline for prompt synthesis and agentic trajectory collection, yielding 23K supervised trajectories and 14.7K RL tasks with three-layer verifiable checklists. We further introduce WeBench-MMGenEdit, a bilingual benchmark covering both knowledge-intensive image generation and multi-image editing. Finally, a two-sided post-training recipe based on SFT and RL improves the agent policy and image backend. Together, WeAgent-MMGenEdit enables a 30B-total/3B-active policy to outperform similarly sized policy models and approach the performance of a 1T-parameter agent.