Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision

Long Cui; Xiaoqian Liu; Qi Qin; Yi Xin; Tao Lin; Jianguo Li; Linfeng Zhang

cs.CVAug 17, 2026

Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision

Authors: Long Cui, Xiaoqian Liu, Qi Qin, Yi Xin, Tao Lin, Jianguo Li, Linfeng Zhang

Organizations: Shanghai Jiao Tong University · Ant Group

Abstract

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.

Figures & tables

Figure 1: (a) Edit Concept Scaling. Left: The previous paradigm is restricted by coarse categories and limited diversity. Right: Our approach scales up to 1,000+ fine-grained concepts to ensure a balanced and rich distribution. (b) Dense Supervision. Left: Conventional training relies on single edit pairs with sparse supervision signals. Right: Our composite edit strategy provides dense supervision, enhancing training efficiency.

Figure 2: Hierarchical taxonomy of 1,000+ fine-grained edit concepts. Top: Overall hierarchical framework. Mid: Specific leaf nodes for detailed edit concepts. Bottom: Visualizations of edit samples.

Figure 3: Overview of the improved synthesis framework. Stage 1: Library Construction leveraging LLM world knowledge. Stage 2: Semantic Matching and Instruction Generation including VQA checklists. Stage 3: Image Synthesis using various editing models. Stage 4: Instance Specific Verification using VQA.

Panel of Figure 4 — Figure 4: Edit concept distributions. Stochastic sampling collapses while our library ensures diversity.

Scale	Method	Fine-Grained Editing Categories									Overall
Scale	Method	Ext.	Add	Style	Rm.	Bg.	Adj.	Rep.	Act.	Comp.	Overall
2M	UnicEdit	2.20	3.71	3.62	1.52	3.48	3.24	3.24	3.46	2.07	2.95
	ScaleEdit	2.10	3.71	4.35	3.22	2.86	2.70	3.57	3.58	2.42	3.17
	ConceptEdit 10	2.18	3.82	3.61	2.00	3.51	3.27	2.69	3.83	2.53	3.05
	ConceptEdit 500	2.19	4.02	4.49	2.46	3.58	3.66	2.43	3.78	2.75	3.26
	ConceptEdit 1000	2.17	3.96	4.55	2.71	3.49	3.59	3.33	3.63	2.52	3.33

Table 1: Quantitative comparison and ablation study on the ImgEdit benchmark [ Ye et al., 2025b ] . The models are trained on 2M and 5M data scales respectively. Abbreviations: Ext.: Extract, Rm.: Remove, Bg.: Background, Adj.: Adjust, Rep.: Replace, Act.: Action, Comp.: Compose.

\Delta

Comp. Gain indicates the improvement brought by dense supervision (w/ Comp) over ConceptEdit 1000 .

\Delta

Overall Gain highlights the total performance margin of our full framework over previous best-performing baseline (ScaleEdit). Best results per category at each scale are in bold.

Figure 5: Training efficiency for different strategies.

Scale	Method	GEdit-Bench-EN			GEdit-Bench-CN
Scale	Method	$G_{SC}\uparrow$	$G_{PQ}\uparrow$	$G_{O}\uparrow$	$G_{SC}\uparrow$	$G_{PQ}\uparrow$	$G_{O}\uparrow$
2M	UnicEdit	4.87	7.03	4.79	4.83	7.00	4.65
	ScaleEdit	5.30	6.79	5.38	5.33	6.66	5.25
	ConceptEdit 10	5.34	6.76	5.41	5.54	7.03	5.15
	ConceptEdit 500	5.90	7.13	5.54	5.51	6.70	5.45
	ConceptEdit 1000	5.91	6.76	5.48	5.87	7.00	5.48

Table 2: Quantitative comparison and ablation study on GEdit-Bench [ Liu et al., 2025 ] . The models are evaluated at 2M and 5M training data scales.

\Delta

Comp. Gain indicates the improvement brought by dense supervision (w/ Comp) over the baseline ConceptEdit 1000 .

\Delta

Overall Gain highlights the performance margin of our full framework over the previous best baseline (ScaleEdit). Best results per metric at each scale are in bold.

Filtering Strategy	Precision	Recall	F1-Score	Accuracy
Filtering Strategy	(%)	(%)	(%)	(%)
Pseudo GT (Gemini)	100.0	100.0	100.0	100.0
Generic Validation	75.0	57.0	65.0	90.0
Instance-Specific (Ours)	84.0	87.0	86.0	95.0
$\Delta$ Gain	$\uparrow$ 9.0	$\uparrow$ 30.0	$\uparrow$ 21.0	$\uparrow$ 5.0

Table 3: Ablation study of our VQA filtering pipeline.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Visualizations of generalized image-to-image translation under various structural control signals.

Metric	Ground Truth	Generic	Ours	$\Delta$ Gain
TN	833	802	805	–
FN	0	72	21	–
TP	167	95	146	–
FP	0	31	28	–
Precision (%)	100.0	75.0	84.0	$\uparrow$ 9.0
Recall (%)	100.0	57.0	87.0	$\uparrow$ 30.0

Appendix

Table 4: Detailed ablation of our VQA filtering pipeline, including raw confusion matrix counts.

Stage	Generic	Ours	Overhead
Instruction Generation	0.523s	0.567s	+0.044s
Filtering Stage	0.312s	0.337s	+0.025s
Total	0.835s	0.904s	+0.069s

Appendix

Table 5: Per-sample processing latency (seconds) benchmarked on Qwen3.5-122B-A10B.

Dataset	Sub- tasks	Size	Post Verification			Basic Edit	Complex Edit	Dist. Control
Dataset	Sub- tasks	Size	Failed Filtration	Inst.-Spec. Filtering	Recaption	Basic Edit	Complex Edit	Dist. Control
MagicBrush [ Zhang et al., 2023 ]	5	$\sim$ 10K	✔	✘	✘	✔	✘	✘
InstructPix2Pix [ Brooks et al., 2023 ]	4	$\sim$ 313K	✘	✘	✘	✔	✘	✘
HQ-Edit [ Hui et al., 2024 ]	6	$\sim$ 197K	✘	✘	✔	✔	✘	✘
SEED-Data-Edit [ Ge et al., 2024 ]	6	$\sim$ 3.7M	✔	✘	✘	✔	✘	✘
UltraEdit [ Zhao et al., 2024 ]	9	$\sim$ 4M	✘	✘	✘	✔	✘	✘

Appendix

Table 6: Comparison of Image Editing datasets.

Model	Advanced	Object	Compo.	Global	Portrait	Text	Overall
Open-source Models
BAGEL-7B-MoT [ Deng et al., 2025 ]	42.21	46.12	38.85	44.90	41.10	33.91	41.66
FLUX.1-Kontext-dev [ Labs et al., 2025 ]	50.73	42.88	44.45	49.05	37.36	40.15	43.53
LongCat-Image-Edit [ Team et al., 2025a ]	64.06	61.46	57.51	65.39	51.68	62.08	59.05
Qwen-Image-Edit-2509 [ Wu et al., 2025 ]	67.36	63.83	58.37	64.59	54.27	69.85	61.27
Qwen-Image-Edit-2511 [ Wu et al., 2025 ]	66.78	68.45	60.38	70.39	53.75	73.50	63.29

Appendix

Table 7: Quantitative comparison on ConceptEdit-Bench. The “Overall” column represents the final aggregated performance across our 1,000 fine-grained categories.

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 302 papers

ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework
Mar 21, 2026Guanzhou Chen, Erfei Cui, Changyao Tian +6Synthetic Data Augmentation Text-Guided Image Editing
Shanghai Jiao Tong University · Shanghai AI Laboratory · CUHK MMLab +3
Training-Free Multi-Concept Image Editing
Feb 24, 2026Niki Foteinopoulou, Ignas Budvytis, Stephan LiwickiTraining-Free Image Editing Image Editing
Cambridge Research Laboratory, Toshiba Europe, Cambridge, UK · Independent Researcher, UK
GIDE: Unlocking Diffusion LLMs for Precise Training-Free Image Editing
Mar 22, 2026Zifeng Zhu, Jiaming Han, Jiaxiang Zhao +2Training-Free Image Editing Image Editing
Xi’an Jiaotong University · MMLab, The Chinese University of Hong Kong

Mar 21, 2026J/K move · Enter open · S save

ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework

Guanzhou Chen, Erfei Cui, Changyao Tian, Danni Yang+5

Shanghai Jiao Tong University · Shanghai AI Laboratory · CUHK MMLab+3

Synthetic Data Augmentation Text-Guided Image Editing Multi-Agent Systems

Open paper AskPDF arXiv

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.

Abstract

Image Editing

Explore similar work

ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework

Training-Free Multi-Concept Image Editing

GIDE: Unlocking Diffusion LLMs for Precise Training-Free Image Editing