Diffusion-based image editing models have achieved significant progress in real world applications. However, conventional models typically rely on natural language prompts, which often lack the precision required to localize target objects. Consequently, these models struggle to maintain background consistency due to their global image regeneration paradigm. Recognizing that visual cues provide an intuitive means for users to highlight specific areas of interest, we utilize bounding boxes as guidance to explicitly define the editing target. This approach ensures that the diffusion model can accurately localize the target while preserving background consistency. To achieve this, we propose FineEdit, a multi-level bounding box injection method that enables the model to utilize spatial conditions more effectively. To support this high precision guidance, we present FineEdit-1.2M, a large scale, fine-grained dataset comprising 1.2 million image editing pairs with precise bounding box annotations. Furthermore, we construct a comprehensive benchmark, termed FineEdit-Bench, which includes 1,000 images across 10 subjects to effectively evaluate region based editing capabilities. Evaluations on FineEdit-Bench demonstrate that our model significantly outperforms state-of-the-art open-source models (e.g., Qwen-Image-Edit and LongCat-Image-Edit) in instruction compliance and background preservation. Further assessments on open benchmarks (GEdit and ImgEdit Bench) confirm its superior generalization and robustness.
Figures & tables
Figure 2: Overview of FineEdit framework, which includes two training stages: (a) Pre-training stage establishes multi-level spatial priors using early and deep fusion. (b) Post-training stage applies reinforcement learning with a novel decoupled reward function.
Figure 3: Unified visual instructions for diverse editing settings. Three representative configurations are showcased: localized style transfer, single-bbox composite editing, and multi-bbox composite editing. FineEdit provides a flexible and consistent interface for both single-target and multi-target manipulation.
Figure 4: The U-GAF Pipeline. , which consists of four synergistic stages for high-quality data synthesis: Data Curation, (2) Data Annotation, (3) Edit Generation, and (4) Data Refinement.
Figure 5: Information of our purposed FineEdit dataset. (a) Training dataset distribution. (b) Evaluation dataset distribution. (c) Human evaluation accuracy on different datasets. (d) Distribution of bounding box area ratios.
Background Preservation
BBox Editing Accuracy
Model
PSNR ↑
SSIM ↑
LPIPS ↓
OBR ↑
CLIP ↑
PC ↑
VN ↑
PDI ↑
Flux-fill-dev [ 21 ]
34.2
0.90
0.08
4.99
0.26
3.56
4.03
3.98
BrushNet [ 17 ]
29.6
0.83
0.13
4.99
0.25
3.28
3.38
3.34
BrushEdit [ 24 ]
30.9
0.84
0.13
4.95
0.25
3.27
3.45
3.41
PrefPaint [ 28 ]
27.1
0.77
0.21
4.83
0.23
2.74
3.01
2.96
Asuka-flux [ 52 ]
34.4
0.88
0.16
5.00
0.22
2.20
2.92
2.85
Table 1: Quantitative Comparison with State-of-the-Art methods on FineEdit-1k Evaluation Bench. PC, VN, PDI, and OBR mean Prompt Compliance, Visual Naturalness, Physical & Detail Integrity, and Out-of-Box Retention, respectively.
Model
Add
Adjust
Extract
Replace
Remove
Background
Style
Hybrid
Action
Overall
OmniGen [ 58 ]
3.47
3.04
1.71
2.94
2.43
3.21
4.19
2.24
3.38
2.96
ICEdit [ 66 ]
3.58
3.39
1.73
3.15
2.93
3.08
3.84
2.04
3.68
3.05
Step1X-Edit [ 29 ]
3.88
3.14
1.76
3.40
2.41
3.16
4.63
2.64
2.52
3.06
BAGEL [ 8 ]
3.56
3.31
1.70
3.30
2.62
3.24
4.49
2.38
4.17
3.20
UniWorld-V1 [ 26 ]
3.82
3.64
2.27
3.47
3.24
2.99
4.21
2.96
2.74
3.26
OmniGen2 [ 57 ]
3.57
3.06
1.77
3.74
3.20
3.57
4.81
2.52
4.68
3.44
Table 2: Quantitative Comparison with SOTA methods on ImgEdit-Bench.
Model
G_SC
G_PQ
G_O
Model
G_SC
G_PQ
G_O
Instruct-P2P [ 5 ]
3.58
5.49
3.68
GPT-Image-1 [High]
7.85
7.62
7.53
AnyEdit [ 62 ]
3.18
5.82
3.21
Gemini 2.0
6.73
6.61
6.32
MagicBrush [ 64 ]
4.68
5.66
4.52
Kontext-dev [ 23 ]
6.52
7.38
6.00
UniWorld-V1 [ 26 ]
4.93
7.43
4.85
Kontext-pro [ 23 ]
7.02
7.60
6.56
OmniGen [ 58 ]
5.96
5.89
5.06
UniPic2 [ 55 ]
-
-
7.10
OmniGen2 [ 57 ]
7.16
6.77
6.41
Qwen-Image-Edit [ 56 ]
8.00
7.86
7.56
Table 3: Quantitative Comparison with SOTA methods on GEdit-Bench-EN.
Figure 6: Comparision on FineEdit-1k Evaluation Bench.
Ablation
Module
Background Preservation
Bbox Editing Accuracy
PSNR ↑
SSIM ↑
LPIPS ↓
OBR ↑
CLIP ↑
PC ↑
VN ↑
PDI ↑
Fusion
Baseline [ 56 ]
27.1
0.80
0.13
4.15
0.26
4.56
4.66
4.65
Early
15.9
0.48
0.40
2.43
0.21
2.45
1.85
1.85
Deep
33.1
0.88
0.10
4.61
0.26
4.52
4.51
4.61
Early + Deep
34.4
0.91
0.08
4.80
0.27
4.67
4.71
4.69
Reward
Pre-RL
34.4
0.91
0.08
4.80
0.27
4.67
4.71
4.69
Table 4: Ablation study on fusion strategies and reward functions.
Figure 7: Extended Qualitative Results on FineEdit-1k
Figure 8: Qualitative Results Across Diverse Editing Tasks
Figure 9: Comparison of human evaluation win rates between FineEdit and competing methods.
Background Preservation
BBox Editing Accuracy
Model
PSNR ↑
SSIM ↑
LPIPS ↓
OBR ↑
CLIP ↑
PC ↑
VN ↑
PDI ↑
FireRed-Image-Edit [ 49 ]
24.2
0.80
0.16
4.06
0.35
4.48
4.52
4.51
Bagel [ 8 ]
29.3
0.87
0.14
4.11
0.35
4.06
3.95
3.93
Stepx-edit-v1p2 [ 29 ]
27.9
0.83
0.13
3.98
0.36
4.47
4.50
4.50
OmniGen2 [ 57 ]
25.1
0.83
0.19
3.32
0.34
3.60
3.69
3.67
FineEdit
34.4
0.91
0.08
4.80
0.27
4.67
4.71
4.69
Table 5: Quantitative Comparison with State-of-the-Art methods. PC, VN, PDI, and OBR mean Prompt Compliance, Visual Naturalness, Physical & Detail Integrity, and Out-of-Box Retention, respectively.
Figure 10: Robustness to Various Bounding Box Scales and Ratios.
Figure 11: Robustness to Bounding Box Perturbations.
City University of Hong Kong (Dongguan), Guangdong, China · City University of Hong Kong, Hong Kong, China · Mohamed bin Zayed University of Artificial Intelligence, Masdar, Abu Dhabi