EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing

Zitong Xu; Huiyu Duan; Zhongpeng Ji; Xinyun Zhang; Yutao Liu; Xiongkuo Min; Ke Gu; Jian Zhang

cs.CVMar 16, 2026

EditHF-1M: A Million-Scale Rich Human Preference Feedback for Image Editing

Authors: Zitong Xu, Huiyu Duan, Zhongpeng Ji, Xinyun Zhang, Yutao Liu, Xiongkuo Min, Ke Gu, Jian Zhang, +4 more

Organizations: Shanghai Jiao Tong University · Vivo Mobile Communication Co., Ltd · University of Electronic and Science Technology of China · Ocean University of China · Beijing University of Technology

Abstract

Recent text-guided image editing (TIE) models have achieved remarkable progress, while many edited images still suffer from issues such as artifacts, unexpected editings, unaesthetic contents. Although some benchmarks and methods have been proposed for evaluating edited images, scalable evaluation models are still lacking, which limits the development of human feedback reward models for image editing. To address the challenges, we first introduce EditHF-1M, a million-scale image editing dataset with over 29M human preference pairs and 148K human mean opinion ratings, both evaluated from three dimensions, i.e., visual quality, instruction alignment, and attribute preservation. Based on EditHF-1M, we propose EditHF, a multimodal large language model (MLLM) based evaluation model, to provide human-aligned feedback from image editing. Finally, we introduce EditHF-Reward, which utilizes EditHF as the reward signal to optimize the text-guided image editing models through reinforcement learning. Extensive experiments show that EditHF achieves superior alignment with human preferences and demonstrates strong generalization on other datasets. Furthermore, we fine-tune the Qwen-Image-Edit using EditHF-Reward, achieving significant performance improvements, which demonstrates the ability of EditHF to serve as a reward model to scale-up the image editing. Both the dataset and code will be released in our GitHub repository: https://github.com/IntMeGroup/EditHF.

Figures & tables

Figure 1: Overview of EditHF-1M and human annotations: (a) Ranking: edited images from different editing models are grouped by source image and editing prompt and ranked to assess relative quality; (b) Scoring: each edited image is rated individually for evaluating absolute quality.

Databases	Source	Edited	Annotations	Editing	Editing	Prompt	Dimensions	Evaluation	Validated for
Databases	Images	Images	Annotations	Models	Tasks	Type	Dimensions	Model	Refinement
TedBench [ 23 ]	100	400	300 Pairs	3	4	Description	Overall	✗	✗
EditVal [ 2 ]	648	5,184	10.3K Scores	8	12	Instruction	Alignment, Preservation	✗	✗
EditEval [ 17 ]	150	1,050	4.2K Scores	7	19	Description, Instruction	Quality, Alignment, Preservation, Realism	✓	✗
AURORA-Bench [ 26 ]	400	2,000	2K Scores	5	8	Instruction	Overall	✓	✗
ImagenHub [ 27 ]	179	1,432	2.8K Scores	8	-	Description, Instruction	Quality, Preservation	✗	✗

Table 1: Comparison of text-guided image editing model evaluation benchmarks. ✓ means publicly available, ✗ means not applicable or not available.

Figure 2: Comparison of image editing models using EditHF-1M. Ranking scores are derived from win counts in pairwise group comparisons.

Figure 3: Comparison of image editing models across different editing tasks.

Figure 4: Overview of the EditHF architecture and the editing model refinement process guided by EditHF.

Dimension	Visual Quality				Editing Alignment				Attribute Preservation
Methods/Metrics	$\mathrm{SRCC}_{global}$	$\mathrm{PLCC}_{global}$	$\mathrm{SRCC}_{group}$	Acc	$\mathrm{SRCC}_{global}$	$\mathrm{PLCC}_{global}$	$\mathrm{SRCC}_{group}$	Acc	$\mathrm{SRCC}_{global}$	$\mathrm{PLCC}_{global}$	$\mathrm{SRCC}_{group}$	Acc
$\heartsuit$ DIIVINE [ 49 ]	0.2062	0.3162	0.1795	0.5633	0.0618	0.1947	0.0962	0.5332	0.1641	0.2686	0.1506	0.5522
$\heartsuit$ BRISQUE [ 46 ]	0.2446	0.2996	0.2020	0.4286	0.0494	0.0975	0.0495	0.4827	0.2019	0.2457	0.1654	0.4426
$\heartsuit$ NIQE [ 47 ]	0.2816	0.3145	0.2386	0.4142	0.0635	0.1432	0.0811	0.4723	0.2623	0.2778	0.2405	0.4157
$\spadesuit$ MSE	0.1421	0.1418	0.2900	0.3968	0.0600	0.0601	0.1730	0.4412	0.3108	0.1893	0.4736	0.3268
$\spadesuit$ PSNR	0.1270	0.0232	0.2844	0.6011	0.0572	0.0134	0.1694	0.5575	0.2978	0.1851	0.4690	0.6708

Table 2: Performance comparisons of quality evaluation methods on EditHF-1M from perspectives of Visual Quality , Editing Alignment , and Attribute Preservation .

\mathrm{SRCC}_{global}

and

\mathrm{PLCC}_{global}

are reported to measure correlation over the entire dataset.

\mathrm{SRCC}_{group}

denotes the average SRCC computed across all image groups. Acc represents the accuracy of pairwise comparison prediction.

\spadesuit

Traditional FR IQA metrics,

\heartsuit

traditional NR IQA metrics,

\clubsuit

deep learning-based FR IQA methods,

\diamondsuit

deep learning-based NR IQA methods, ★ vision-language methods, ✩ MLLM-based models. The fine-tuned results are marked with ✻ . The best results are highlighted in red , and the second-best results are highlighted in blue .

Dimensions	Perceptual Quality				Editing Alignment				Attribute Preservation				Overall Score				Overall Rank
Models/Metrics	Human	Ours ✻	Q-align ✻	EditReward	Human	Ours ✻	Qwen3-VL ✻	EditReward	Human	Ours ✻	AHIQ ✻	EditScore	Human	Ours ✻	EditScore	EditReward	Human	Ours
NanoBanana [ 7 ]	17.54	17.61	16.20	5.863	17.01	17.97	17.24	6.746	18.56	18.43	18.65	18.85	17.68	17.70	5.608	1.810	1	1
Qwen-Image-Edit [ 66 ]	15.26	16.53	14.51	6.453	18.17	18.26	17.38	6.952	17.51	17.86	19.95	18.19	17.15	17.33	5.463	1.799	2	2
Seedream4 [ 56 ]	15.94	16.77	14.85	6.377	17.31	18.06	17.31	7.013	16.57	17.38	19.34	17.85	16.74	17.19	5.352	1.799	3	4
FLUX-Kontext [ 29 ]	15.12	16.37	14.87	6.602	17.12	18.04	17.13	7.126	17.40	18.06	20.13	17.05	16.65	17.25	5.259	1.794	4	3
DreamOmni [ 71 ]	14.33	15.91	13.57	6.658	16.14	17.45	16.69	7.109	18.08	18.22	20.06	16.59	16.19	16.93	5.234	1.794	5	5

Table 3: Comparisons of the alignment between different evaluation methods and human perception in evaluating image editing models. The best results are highlighted in red , and the second-best results are highlighted in blue . ✻ denotes fine-tuned models.

Backbone $\&$ Strategy				Perceptual Quality			Editing Alignment			Attribute Preservation
Backbone	MLP Head	Pointwise Training	Pairwise Training	$\mathrm{SRCC}_{global}$	$\mathrm{SRCC}_{group}$	Acc	$\mathrm{SRCC}_{global}$	$\mathrm{SRCC}_{group}$	Acc	$\mathrm{SRCC}_{global}$	$\mathrm{SRCC}_{group}$	Acc
InternVL3.5 [ 63 ]		✓		0.7184	0.7215	0.7904	0.7025	0.7312	0.7750	0.7241	0.7506	0.7862
InternVL3.5 [ 63 ]	✓	✓		0.7762	0.7829	0.7826	0.7806	0.7964	0.7825	0.7752	0.7864	0.7821
InternVL3.5 [ 63 ]	✓		✓	0.8018	0.8792	0.8663	0.8213	0.8894	0.8762	0.8173	0.8726	0.8765
InternVL3.5 [ 63 ]	✓	✓	✓	0.8548	0.9223	0.9152	0.8488	0.9319	0.9298	0.8475	0.9251	0.9197
InternVL3 [ 5 ]	✓	✓	✓	0.8316	0.8964	0.8826	0.8155	0.9109	0.9035	0.8141	0.9022	0.8943

Table 4: Ablation study on the different backbones and training strategy.

Methods/Benchmarks	GenAI-Bench [ 20 ]	AROURA-Bench [ 26 ]	ImagenHub [ 27 ]	EBench-18K [ 74 ]	EditScore-Bench [ 41 ]	EditReward-Bench [ 69 ]
GPT-4o [ 51 ]	0.5354	0.5081	0.3821	0.4108	0.6730	0.2831
GPT-5 [ 52 ]	0.5961	0.4727	0.4085	0.4260	0.7524	0.3781
Gemini-2.0-Flash [ 13 ]	0.5332	0.4431	0.2369	0.3892	0.6958	0.3347
Gemini-2.5-Flash [ 7 ]	0.5701	0.4763	0.4162	0.4337	0.7026	0.3802
Qwen3-VL (7B) [ 77 ]	0.4458	0.3428	0.2426	0.3726	0.4251	0.2731
InternVL3.5 (7B) [ 63 ]	0.4517	0.3692	0.2071	0.3511	0.4518	0.2406

Table 5: Comparison results on other public image editing benchmarks. The best results are highlighted in red , and the second-best results are highlighted in blue . “-” indicates that the model was trained on this dataset, and thus its result is not reported to ensure a fair cross-dataset evaluation.

Benchmarks	EditHF-1M		AROURA-Bench [ 26 ]		ImagenHub [ 27 ]		EBench-18K [ 74 ]		EditScore-Bench [ 41 ]		EditReward-Bench [ 69 ]
Methods/Metrics	Human	Ours	Human	Ours	Human	Ours	Human	Ours	Human	Ours	Human	Ours
Qwen-Image-Edit	75.39	78.06	83.21	68.47	75.64	73.26	73.25	78.29	70.17	75.22	80.15	77.03
Qwen-Image-Edit (Self-DPO)	78.36	81.55	84.28	69.26	76.55	75.14	73.80	79.05	72.18	76.55	81.17	78.46
Improvement	+3.94%	+4.47%	+1.29%	+1.15%	+1.20%	+2.57%	+0.75%	+0.97%	+2.86%	+1.77%	+1.27%	+1.86%
Qwen-Image-Edit (Global-DPO)	84.26	86.15	85.46	69.12	79.12	78.06	74.18	79.60	75.23	78.41	82.60	78.12
Improvement	+11.77%	+10.36%	+2.70%	+0.95%	+4.60%	+6.55%	+1.27%	+1.67%	+7.21%	+4.24%	+3.06%	+1.42%

Table 6: Results of Qwen-Image-Edit [ 66 ] refined with our EditHF on different benchmarks. Overall scores from human study and our EditHF are reported.

Figure 5: Examples from Qwen-Image-Edit refined by our EditHF and other competitive editing models.

Explore similar work

ReasonEdit: Towards Interpretable Image Editing Evaluation via Reinforcement Learning
May 8, 2026Honghua Chen, Zitong Xu, Huiyu Duan +3Image Editing Evaluation Text-Guided Image Editing
University of Electronic Science and Technology of China · 2Shanghai Jiao Tong University
HP-Edit: A Human-Preference Post-Training Framework for Image Editing
Apr 21, 2026Fan Li, Chonghuinan Wang, Lina Lei +9Diffusion Model Alignment Image Editing Evaluation
Huawei Noah’s Ark Lab · Nankai University · Harbin Institute of Technology
EditRefiner: A Human-Aligned Agentic Framework for Image Editing Refinement
May 8, 2026Zitong Xu, Huiyu Duan, Yifei Nie +9Text-Guided Image Editing Image Editing
Shanghai Jiao Tong University · University of Electronic Science and Technology of China · Vivo Mobile Communication Co., Ltd

May 8, 2026J/K move · Enter open · S save

ReasonEdit: Towards Interpretable Image Editing Evaluation via Reinforcement Learning

Honghua Chen, Zitong Xu, Huiyu Duan, Xinyun Zhang+2

University of Electronic Science and Technology of China · 2Shanghai Jiao Tong University

Image Editing Evaluation Text-Guided Image Editing Multimodal Reward Modeling

Open paper AskPDF arXiv

Recent text-guided image editing (TIE) models have achieved remarkable progress, however, many edited results still suffer from artifacts, unintended modifications, and suboptimal aesthetics. Although several benchmarks and evaluation methods have been proposed, most existing approaches rely on scalar scores and lack interpretability. This limitation largely stems from the absence of high-quality interpretation datasets for TIE and effective reward models to train interpretable evaluators. To address these challenges, we introduce ReasonEdit-22K, the first dataset that combines 22K edited images with 113K Chain-of-Thought (CoT) samples, along with 1.3M human judgments assessing these interpretations in terms of logicality, accuracy, and usefulness. Building upon this dataset, we propose RE-Reward, a multimodal large language model (MLLM)-based reward model designed to provide human-aligned feedback for evaluating interpretable reasoning in image editing. Furthermore, we develop ReasonEdit, which is trained using reward signals derived from RE-Reward and the Group Relative Policy Optimization (GRPO) algorithm to learn an interpretable evaluation model. Extensive experiments demonstrate that ReasonEdit achieves superior alignment with human preferences and exhibits strong generalization across public benchmarks. In addition, it is capable of generating high-quality interpretable evaluation text, enabling more transparent and trustworthy assessment for image editing. The code is available at https://github.com/IntMeGroup/ReasonEdit.

Abstract

Explore similar work

ReasonEdit: Towards Interpretable Image Editing Evaluation via Reinforcement Learning

HP-Edit: A Human-Preference Post-Training Framework for Image Editing

EditRefiner: A Human-Aligned Agentic Framework for Image Editing Refinement