Commercial and advertising images are frequently affected by poor framing, partially cropped subjects, truncated text or logos, and insufficient context, all of which can reduce subject clarity, i.e., the ability of an image to clearly communicate its primary subject. Image outpainting offers a scalable solution by extending image boundaries and recovering missing content and context. However, existing diffusion-based outpainting methods often produce visually plausible completions while degrading subject fidelity through structural inconsistencies, semantic drift, or loss of fine-grained detail. To address this limitation, we propose a subject clarity outpainting framework that combines vision-language model (VLM)-guided semantic conditioning with multiscale wavelet supervision for subject-localized detail preservation. To support training, we develop a subject-centric data curation pipeline that constructs subject-intersecting outpainting pairs from advertising and natural images. The resulting objective introduces no additional inference cost and is designed to be compatible with diffusion-based backbones. Across four advertising and natural-image benchmarks, our method improves subject clarity, reducing subject-centered DreamSim error and FID on average by 3.0% and 2.4% over matched supervised fine-tuning, and by 10.8% and 7.7% over the strongest state-of-the-art approach per dataset, respectively.
Figures & tables
Figure 1: What is subject clarity outpainting? General outpainting extends an image beyond its boundaries to synthesize missing scene content. We focus on the regime in which the original framing partially crops the primary subject. In this setting, outpainting must reconstruct the missing subject region while preserving subject fidelity —identity, structure, fine-grained appearance, and continuity with the visible portion. Existing methods can generate plausible surroundings yet distort the continued subject.
Figure 2: Overview of the proposed training framework. vision–language model provides scene guidance to the outpainting backbone. During training, the predicted and ground-truth images are compared in a multilevel wavelet space within the affected subject region. Subject-frequency statistics determine the relative importance of wavelet subbands, while the noise level controls the active scales. The resulting subject-aware wavelet loss complements the standard flow objective and is not used during inference.
Method
Backbone
Synth-Ads
Real-Ads
Overall
Subject Clarity
Overall
Subject Clarity
FID ↓
LPIPS full ↓
LPIPS Ω ↓
DINOv2 Ω ↓
DreamSim ctx ↓
FID ↓
LPIPS full ↓
LPIPS Ω ↓
DINOv2 Ω ↓
DreamSim ctx ↓
BrushNet [ 9 ]
SDXL
31.41
0.2590
0.5015
0.4586
0.0743
74.90
0.2839
0.5528
0.5124
0.1096
PowerPaint [ 32 ]
SD 1.5
31.79
0.2590
0.4152
0.3789
0.0815
72.77
0.2741
0.4799
0.4453
0.1184
VIP [ 29 ]
SD 1.5
25.26
0.2118
0.3957
0.3490
0.0514
55.20
0.2113
0.4064
0.3593
0.0783
FLUX.1 Fill [ 11 ]
FLUX 1.0
29.56
0.2327
0.3542
0.3157
0.0644
86.92
0.3148
0.4545
0.4083
0.1555
Table 1: Outpainting results on the Synth-Ads and Real-Ads test sets. Per-dataset best and second best are bold and underlined , respectively.
Method
Backbone
LVIS
Open Images
Overall
Subject Clarity
Overall
Subject Clarity
FID ↓
LPIPS full ↓
LPIPS Ω ↓
DINOv2 Ω ↓
DreamSim ctx ↓
FID ↓
LPIPS full ↓
LPIPS Ω ↓
DINOv2 Ω ↓
DreamSim ctx ↓
BrushNet [ 9 ]
SDXL
19.91
0.2587
0.5173
0.4910
0.0991
60.41
0.2674
0.5159
0.5100
0.1102
PowerPaint [ 32 ]
SD 1.5
22.49
0.2788
0.4530
0.4262
0.1065
66.27
0.2853
0.4588
0.4520
0.1193
VIP [ 29 ]
SD 1.5
15.82
0.2082
0.4107
0.3615
0.0708
48.18
0.2116
0.4062
0.3628
0.0756
FLUX.1 Fill [ 11 ]
FLUX 1.0
22.76
0.2594
0.3953
0.3510
0.0951
63.10
0.2716
0.3929
0.3686
0.1020
Table 2: Outpainting results on the LVIS and Open Images test sets. Per-dataset best and second best are bold and underlined , respectively.
Figure 3: The Input / GT column shows the model input and ground-truth outpaint region in gray, with the black line marking the outpaint boundary. The prompts used for these examples are reported in Tab. 10 . Relative to representative external baselines and the matched SFT model, our method more consistently preserves subject identity and produces coherent continuations across the boundary.
Dataset
Method
FID
LPIPS full
LPIPS Ω
DINOv2 Ω
DreamSim ctx
Synth-Ads
FLUX.2 SFT
24.07
0.2103
0.3524
0.3134
0.0479
GPT-Image-2
23.59
0.2083
0.4274
0.3816
0.0454
Ours
23.88
0.2094
0.3539
0.3154
0.0475
Real-Ads
FLUX.2 SFT
49.25
0.2010
0.3754
0.3200
0.0660
GPT-Image-2
50.19
0.2135
0.4828
0.4241
0.0677
Ours
47.31
0.1999
0.3686
0.3156
0.0629
Table 3: Comparison on test samples with available GPT-Image-2 predictions (95.1% overall; 97.2% Synth-Ads, 96.3% Real-Ads, 94.1% LVIS, and 93.6% Open Image). Best in bold ; second best underlined . Lower is better.
Figure 4: Comparison with GPT-Image-2. Representative comparisons illustrating (a) subject-continuation misalignment, and (b) failure to preserve subject structure.
Table 5: Component ablation on Real-Ads. Best in bold ; second best underlined . Lower is better.
Figure 5: Representative qualitative ablation. N.S. & S.A. denote noise schedule and subband adaptation, respectively. Additional examples and analysis are provided in Sec. C.5 .
Figure 6: Attention-routing analysis. Direct attention from generated-image queries to the prompt subwords for “foreleg.” SFT spreads this cue broadly into the outpaint region, whereas ours concentrates it on the supported subject continuation. Maps are overlaid on dimmed predictions and span all image queries, whereas the reported metrics use outpaint queries only.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Input
Kept
Yield
Synth-Ads
5,659
3,707
65.5%
Real-Ads
1,356
969
71.5%
LVIS
19,808
6,705
33.8%
Open Images
4,650
1,267
27.2%
Appendix
Table 6: High-level filtering statistics. Yield is the percentage of input images retained as finalized pairs.
Dataset
Train
Val.
Test
Total
Synth-Ads
2,595
178
934
3,707
Real-Ads
681
48
240
969
LVIS
4,677
377
1,651
6,705
Open Images
904
65
298
1,267
Total
8,857
668
3,123
12,648
Appendix
Table 7: Realized train/validation/test split sizes after filtering.
Table 9: Distribution (%) of samples by realized masked fraction of the full image. Rows may differ from 100% by rounding.
Baseline
Prompt(s)
PowerPaint
All rows use the native image-outpainting context token with an empty free-form prompt.
FLUX.1 Fill
(a) dog wearing a black coat, with the dog’s upper body, back, neck, and head continuing from the visible coat. (b) antique handwritten manuscript page and stacked books, with manuscript page with cursive writing and leather-bound books continuing from the visible boundary. (c) seaplane, with the seaplane’s upper fuselage, cockpit windows, wings, tail, and propeller continuing above the floats. (d) giraffe, with the cropped giraffe’s upper neck and head, with surrounding dry savanna brush continuing naturally.
fal/FLUX.2 LoRA
All rows use the checkpoint-specific instruction Fill the green spaces according to the image.
FLUX.2 SFT
(a) Fill the gray region by extending dog wearing a black coat, with the dog’s upper body, back, neck, and head continuing from the visible coat. (b) Fill the gray region by extending antique handwritten manuscript page and stacked books, with manuscript page with cursive writing and leather-bound books continuing from the visible boundary. (c) Fill the gray region by extending seaplane, with the seaplane’s upper fuselage, cockpit windows, wings, tail, and propeller continuing above the floats. (d) Fill the gray region by extending giraffe, with the cropped giraffe’s upper neck and head, with surrounding dry savanna brush continuing naturally.
Ours
Same as FLUX.2 SFT.
Appendix
Table 10: Prompts used for the predictions in the main paper qualitative comparison. Labels (a)–(d) follow the figure rows. Prompt strings are reproduced as used at inference.
Method
Trainable params.
Train complexity
Time/image
FLUX.2 SFT
39.977M
X
13.32 s
Ours
39.977M
X+O(CHW)
13.32 s
Appendix
Table 11: Computational cost relative to matched SFT. Complexity denotes per-step training work. Time/image is the sample-weighted wall-clock mean over the test set at 30 denoising steps.
Figure 8: Metric-selection behavior. Each column shows the prediction ranked best by the indicated metric among the methods; the best and second-best methods and scores appear below the image to expose the selection margin. Higher is better for masked PSNR and SSIM, whereas lower is better for LPIPS Ω and DreamSim ctx .
Figure 10: Example MLLM judgment on LVIS. Ours is preferred over matched SFT for all four reported criteria. In the anonymous evaluation, FLUX.2 SFT was presented as Method A and ours as Method B; the model identities in parentheses are shown only for the reader. The judge saw the masked input, binary mask, and the two predictions, but not the ground-truth completion. The text below the images condenses the criterion-level rationales returned by GPT-5.5.
Figure 11: Auxiliary high-frequency loss-weight sensitivity. FID for fixed global coefficients and a sample-adaptive coefficient. Lower is better.
Figure 12: Two complementary directions of visual attention. Rows compare ours with matched SFT; columns show denoising steps. Both models retrieve evidence from the correct visible subject in (a) , but differ in where they deploy that evidence in (b) .
Visible-subject routing
“Foreleg” routing
Step
μO
R
μO
R
SFT
Ours
SFT
Ours
SFT
Ours
SFT
Ours
5
0.0341
0.0277
2.56
3.23
0.00341
0.00056
0.66
0.31
10
0.0251
0.0207
4.03
5.15
0.00238
0.00102
0.67
0.59
20
0.0162
0.0068
4.91
13.54
0.00209
0.00024
0.90
5.39
30
0.0063
0.0026
5.13
11.63
0.00087
0.00022
1.47
4.91
Appendix
Table 12: Step-resolved routing metrics in the final double-stream block. Lower μO indicates less unsupported-region routing, while higher R indicates better localization. Best values within each probe and step are bold.
Figure 13: Word-conditioned routing. Direct attention from image queries to selected prompt subwords. The “foreleg” cue is the clearest discriminator: SFT spreads it into unsupported locations, consistent with the duplicated limb, whereas ours increasingly confines it to the valid subject continuation
Figure 20: Additional comparisons with GPT-Image-2 (Part 1). Each row is a separate comparison panel; columns show the Input / GT visualization, GPT-Image-2, and our model.
Figure 21: Additional comparisons with GPT-Image-2 (Part 2). Each row is a separate comparison panel; columns show the Input / GT visualization, GPT-Image-2, and our model.
Figure 22: Additional comparisons with GPT-Image-2 (Part 3). Each row is a separate comparison panel; columns show the Input / GT visualization, GPT-Image-2, and our model.
Recent subject-to-image models have achieved impressive progress in personalized image generation, yet they still struggle to preserve fine-grained subject-specific details. A major reason is the lack of high-quality fine-grained identity supervision: real paired data are expensive to collect, while synthesized training pairs often preserve only coarse subject appearance and fail to capture subtle subject-specific details. In this work, we propose CopyCat, a lightweight model-refinement framework that improves fine-grained subject consistency within only a few seconds. CopyCat performs a one-time refinement of a pretrained subject-to-image model by attaching a lightweight Fine-grained Consistency LoRA (FCLoRA) and optimizing it using a single proxy image, which is used as both the conditioning image and the reconstruction target. This exact self-reconstruction objective substantially simplifies the optimization task, enabling effective fine-grained refinement within only a few seconds. The refinement is performed only once; the resulting model can be directly applied to diverse unseen reference subjects and prompts without further subject-specific optimization. We further revisit subject-to-image LoRA training in double-stream diffusion transformers and find that adapting only the visual stream consistently improves subject consistency. Extensive experiments on DreamBench and XVerseBench demonstrate consistent improvements in fine-grained subject consistency across representative subject-to-image models under both single- and multi-subject settings.
Peng Zheng, Ruiqi Liu, Rui Ma +1
School of Artificial Intelligence, Jilin University · 2Shanghai Innovation Institute · 4Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MOE, China +1
We present FlowFixer, a refinement framework for subject-driven generation (SDG) that restores fine details lost during generation caused by changes in scale and perspective of a subject. FlowFixer proposes direct image-to-image translation from visual references, avoiding ambiguities in language prompts. To enable image-to-image training, we introduce a one-step denoising scheme to generate self-supervised training data, which automatically removes high-frequency details while preserving global structure, effectively simulating real-world SDG errors. We further propose a keypoint matching-based metric to properly assess fidelity in details beyond semantic similarities usually measured by CLIP or DINO. Experimental results demonstrate that FlowFixer outperforms state-of-the-art SDG methods in both qualitative and quantitative evaluations, setting a new benchmark for high-fidelity subject-driven generation.
Image outpainting extends an image beyond its original borders, requiring seamless style integration and globally coherent scene completion. Building on the success of diffusion models, recent methods have achieved substantial improvements in visual quality. In practice, however, high-resolution outpainting is commonly performed via progressive expansion around a fixed source image, particularly in artwork scenarios. Despite this progress, existing approaches still suffer from three key limitations: (i) the absence of a reliable global planning mechanism, which leads to structural instability and error accumulation at high resolutions; (ii) limited spatial controllability beyond text prompts, making it difficult to place objects at user-specified locations; and (iii) high inference latency caused by inherently sequential patch generation. To address these issues, we propose a global blueprint-guided two-stage diffusion framework for layout-controllable high-resolution outpainting with efficient parallel synthesis. In Stage 1, we generate a low-resolution global blueprint using a layout adapter that injects bounding-box conditions into a Stable Diffusion inpainting backbone, producing a globally consistent structural plan while extracting global guidance features. In Stage 2, we synthesize high-resolution local patches in parallel by injecting the blueprint-derived global guidance and initializing each patch from the blueprint using the low-frequency preservation property of forward diffusion. This design eliminates sequential dependency while maintaining global coherence. Extensive experiments on large-scale artwork datasets demonstrate improved visual fidelity, stronger semantic consistency, and substantially reduced inference time compared to prior baselines, while uniquely supporting explicit layout control for artwork outpainting.