Conditional image generation methods are increasingly used in human-centric applications, yet existing human amodal completion (HAC) models offer users limited control over the completed content. Given an occluded person image, they hallucinate invisible regions while preserving visible ones, but cannot reliably incorporate user-specified constraints such as a desired pose or spatial extent. As a result, users often resort to repeatedly sampling the model until they obtain a satisfactory output. Pose-guided person image synthesis (PGPIS) methods allow explicit pose conditioning, but frequently fail to preserve the instance-specific visible appearance and tend to be biased toward the training distribution, even when built on strong diffusion model priors. To address these limitations, we introduce promptable human amodal completion (PHAC), a new task that completes occluded human images while satisfying both visible appearance constraints and multiple user prompts. Users provide simple point-based prompts, such as additional joints for the target pose or bounding boxes for desired regions; these prompts are encoded using ControlNet modules specialized for each prompt type. These modules inject the prompt signals into a pre-trained diffusion model, and we fine-tune only the cross-attention blocks to obtain strong prompt alignment without degrading the underlying generative prior. To further preserve visible content, we propose an inpainting-based refinement module that starts from a slightly noised coarse completion, faithfully preserves the visible regions, and ensures seamless blending at occlusion boundaries. Extensive experiments on the HAC and PGPIS benchmarks show that our approach yields more physically plausible and higher-quality completions, while significantly improving prompt alignment compared with existing amodal completion and pose-guided synthesis methods.
Figures & tables
Figure 1: Assume that, as in (A), a single input image is paired with multiple target poses. The desired behavior is that each generated image aligns with its target pose while preserving the visible appearance. However, prior human amodal completion (HAC) methods such as SDHDO [ 41 ] often fail to align with the specified pose, as shown in (B), despite leveraging pose information. Pose-guided person image synthesis (PGPIS) methods such as MCLD [ 30 ] align well with the pose condition but degrade the visible appearance of the input, especially around the shoes, as shown in (C). While SDHDO preserves the visible appearance better than MCLD, its results remain blurry and show noticeable degradation. In contrast, given the multi-condition inputs in (A), our method simultaneously aligns with the target pose and preserves the visible appearance, producing the user-intended human image.
Figure 2: User prompts . Users specify the intended pose with point prompts, which we use to condition the model. For the pose prompt ppo , we use OpenPose [ 5 ] to detect the visible joints, show them to the user, who then adds the missing joints for the desired pose. Alternatively, the user selects two points to specify a bbox prompt, choosing either an interest-region bbox pib or an entire-region bbox peb . To provide fine-grained control, the pose and bbox prompts can be combined, yielding ppoib or ppoeb . To make effective use of the spatial information, we convert the point coordinates into a prompt image Ip and use it as a conditioning input.
Figure 3: Method overview . Given an incomplete image Iic and a user prompt P , our PHAC framework processes them through (A) coarse image generation and (B, C) a refinement stage. In (A), the denoising U-Net ϵcig starts from random noise and denoises it for T steps to generate a coarse complete image Icc , conditioned on a prompt image Ip (see Fig. 2 ); Ip is fed to a prompt-specific ControlNet ΦCN to provide conditioning, and only the cross-attention blocks of ϵcig are fine-tuned to preserve the pre-trained prior. In (B), invisible mask prediction U-Net Uiv predicts an invisible mask Miv , which is then dilated to Miv′ . In (C), we construct the base composite Ibase and add low-magnitude noise to the invisible region only. The refinement network ΦRF then takes the noisy Ibase as input and outputs the refined completion Irc , preserving the visible region while refining the coarse completion and mitigating boundary artifacts.
Method
User Prompt
OccThuman2.0 test dataset (synthetic)
AHP test dataset (real)
LPIPS* ↓
SSIM ↑
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
LPIPS* ↓
SSIM ↑
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
PIDM [ 4 ]
2D pose map
126.33
0.797
56.91
27.34
16.80
113.72
143.15
0.804
23.71
25.33
16.61
40.47
MCLD [ 30 ]
UV map
115.90
0.833
41.11
18.94
18.37
53.38
121.80
0.853
21.90
14.85
18.97
11.53
pix2gestalt [ 42 ]
-
90.11
0.911
16.51
7.58
22.63
36.65
75.73
0.942
5.98
5.22
24.06
10.96
pix2gestalt [ 42 ] †
2D pose map
88.58
0.914
16.75
6.93
22.94
31.37
75.25
0.943
6.35
4.87
24.25
10.42
SDHDO [ 41 ]
2D pose map
81.39
0.924
16.41
7.05
23.80
43.49
64.19
0.956
6.05
6.05
24.45
9.24
Table 1: Quantitative comparison for promptable human amodal completion. We use the same 2D pose map as the user prompt, whereas MCLD uses UV maps following its experimental setting. The best performance for each metric is highlighted in bold , while the second-best is underlined . † indicates injection of the same user prompt via a pre-trained ControlNet for fair comparison. * indicates that values are reported ×103 for readability.
Figure 4: Qualitative comparison on the AHP test dataset. (A) Partial RGB results; (B) Generated images with different seeds. PGPIS baselines (PIDM, MCLD) frequently hallucinate training set appearances. Amodal completion baselines (pix2gestalt, SDHDO) do not preserve the visible appearance and often violate the pose condition. In contrast, our approach yields consistent pose alignment across seeds and preserves the visible regions.
User Prompt
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
Joint Err. ↓
Pose ( ppo )
49.47
0.948
4.37
25.86
23.33
Interest bbox ( pib )
51.83
0.942
5.69
24.99
24.01
Entire bbox ( peb )
52.28
0.941
5.66
25.07
28.23
Pose & Interest bbox ( ppoib )
49.35
0.947
4.56
25.69
22.15
Pose & Entire bbox ( ppoeb )
49.42
0.946
4.89
25.49
21.96
Table 2: Quantitative comparison across user prompts. On OccThuman2.0, the pose prompt ppo yields the best reconstruction metrics, while bbox prompts (pib,peb) perform worse and exhibit higher joint errors. Combining pose and bbox (ppoib,ppoeb) consistently reduces joint error, improving alignment.
s
OccThuman2.0 test dataset
AHP test dataset
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
0.1
49.39
0.948
4.41
25.82
39.02
0.970
3.46
26.84
0.3
49.49
0.948
4.39
25.83
38.97
0.970
3.44
26.88
0.5
49.47
0.948
4.36
25.85
38.77
0.970
3.41
26.93
0.7
50.18
0.947
4.43
25.27
39.03
0.969
3.45
26.81
0.9
51.91
0.943
4.77
25.39
40.02
0.967
3.68
26.33
Table 3: Effect of the refinement noise strength s . Small s yields limited improvement due to minimal denoising, whereas large s degrades the original appearance.
Method
LPIPS* ↓
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
MCLD [ 30 ]
59.56 (49%)
11.77 (71%)
7.50 (60%)
23.36 (27%)
31.41 (41%)
pix2gestalt [ 42 ]
57.03 (37%)
8.14 (51%)
6.21 (18%)
24.39 (27%)
30.65 (16%)
SDHDO [ 41 ]
59.41 (27%)
10.92 (34%)
6.80 (4%)
24.12 (1%)
33.97 (22%)
Ours
49.47 (16%)
6.12 (18%)
4.37 (14%)
25.86 (5%)
23.33 (1%)
Table 4: Plug-and-play application of the refinement network. Blue marks results improved by refinement, and % reports the relative improvement over the baseline.
User Prompt
OccThuman2.0 test dataset (synthetic)
AHP test dataset (real)
LPIPS* ↓
SSIM ↑
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
LPIPS* ↓
SSIM ↑
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
Pose ( ppo )
49.47
0.948
6.12
4.37
25.86
23.33
38.77
0.970
1.25
3.41
26.93
6.37
Interest bbox ( pib )
51.83
0.942
6.89
5.69
24.99
24.01
41.03
0.964
1.44
4.87
25.99
11.18
Entire bbox ( peb )
52.28
0.941
6.32
5.66
25.07
28.23
41.35
0.963
1.37
5.04
25.78
11.84
Pose & Interest bbox ( ppoib )
49.35
0.947
6.03
4.56
25.69
22.15
38.67
0.970
1.31
3.42
26.60
6.21
Pose & Entire bbox ( ppoeb )
49.42
0.946
6.12
4.89
25.49
21.96
38.51
0.971
1.33
3.48
26.74
6.15
Table S5 : Quantitative comparison across different user prompts.
ppo
pib
peb
ppoib
ppoeb
Δ JE
6.43
5.75
1.53
7.61
7.80
Δ JE pp
1.07
2.86
0.77
0.95
0.98
Table S6 : Quantitative comparison of prompt efficiency. We report the absolute gain in joint error ( Δ JE) and the per-point gain ( Δ JE pp) for different user prompts. While pose-based prompts yield larger absolute gains, the interest-region bbox prompt pib achieves the largest per-point gain, indicating efficient improvements with minimal user input.
Figure S6 : Construction pipeline of the OccThuman2.0 dataset. For each THuman2.0 mesh, we render 10 views and apply per-view data augmentations and background-occlusion compositing to generate 10 composited images. Repeating this process across all 526 meshes produces a total of 5,260 composited images.
Figure S7 : Prompt image for MCLD. MCLD uses a DensePose-based UV map as its conditioning input instead of a 2D pose map. We obtain the target UV map from the occlusion-free complete image and perform MCLD inference conditioned on this UV map.
Method
OccThuman2.0
AHP
mIoU ↑
L1* ↓
mIoU ↑
L1* ↓
pix2gestalt [ 42 ]
48.26
39.74
48.97
22.04
SDHDO [ 41 ]
51.86
33.58
51.47
21.35
Ours
87.84
6.04
77.87
7.89
Table S7 : Quantitative results for invisible mask prediction. Previous amodal completion methods [ 42 , 41 ] do not use the coarse complete image Icc as input, so their results are not directly comparable.
Figure S8 : Qualitative results for invisible mask prediction. Given the coarse complete image Icc , Uiv predicts invisible masks Miv that closely match the ground truth.
Figure S9 : Qualitative results for the refinement process. (A) Results on the AHP dataset and (B) results on the OccThuman2.0 dataset. The blue bboxes compare the facial regions of the input image Iic , the coarse complete image Icc , and the refined image Irc to assess preservation of visible appearance. The red bboxes compare the boundary regions of the input image Iic , the base composite image Ibase , and the refined image Irc to verify smooth blending without boundary artifacts.
Method
User Prompt
OccThuman2.0 test dataset (synthetic)
AHP test dataset (real)
LPIPS* ↓
SSIM ↑
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
LPIPS* ↓
SSIM ↑
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
pix2gestalt [ 42 ]
-
90.11
0.911
16.51
7.58
22.63
36.65
75.73
0.942
5.98
5.22
24.06
10.96
pix2gestalt [ 42 ] †
2D pose map
88.58
0.914
16.75
6.93
22.94
31.37
75.25
0.943
6.35
4.87
24.25
10.42
SDHDO [ 41 ]
2D pose map
81.39
0.924
16.41
7.05
23.80
43.49
64.19
0.956
6.05
6.05
24.45
9.24
Ours
-
53.19
0.939
6.23
5.69
25.03
29.76
41.99
0.964
1.60
4.79
25.86
13.23
Ours
2D pose map
49.47
0.948
6.12
4.37
25.86
23.33
38.77
0.970
1.25
3.41
26.93
6.37
Table S8 : Quantitative comparison for PHAC. This table largely mirrors Table 1 of the main paper, but additionally includes the results of our method evaluated without a user prompt. To avoid redundancy, PGPIS results are omitted. † indicates the use of the same user prompt injected through a pre-trained ControlNet.
Method
OccThuman2.0
AHP
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
PIDM [ 4 ]
88.43
0.862
16.16
19.31
120.85
0.835
20.40
17.51
MCLD [ 30 ]
76.77
0.894
11.48
20.73
101.23
0.880
11.82
20.05
pix2gestalt [ 42 ]
48.49
0.976
1.37
29.05
52.81
0.971
1.55
28.55
SDHDO [ 41 ]
38.08
0.996
0.32
35.23
37.71
0.994
0.32
35.27
Ours
13.72
0.998
0.17
37.93
15.65
0.998
0.14
38.85
Table S9 : Quantitative comparison for PHAC in the visible region. We compute LPIPS, SSIM, MSE, and PSNR on the visible (non-occluded) regions of the input image to measure fidelity to the input.
Method
OccThuman2.0
AHP
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
LPIPS* ↓
SSIM ↑
MSE* ↓
PSNR ↑
PIDM [ 4 ]
51.99
0.908
11.50
21.37
27.44
0.966
4.92
26.03
MCLD [ 30 ]
47.67
0.927
7.65
23.25
23.94
0.974
3.34
27.24
pix2gestalt [ 42 ]
45.79
0.932
6.36
24.32
23.96
0.972
3.67
27.43
SDHDO [ 41 ]
47.14
0.926
6.90
24.22
27.57
0.960
5.73
25.39
Ours
36.85
0.950
4.29
26.15
23.05
0.973
3.28
27.60
Table S10 : Quantitative comparison for PHAC in the invisible region. Using the same metrics as Table S9 , we compute them on the invisible (occluded) regions to measure reconstruction fidelity, conditioned on the visible appearance of the input image.
Figure S10 : Additional qualitative comparison for PHAC. (A, B): AHP test examples, where previous methods either hallucinate appearance, lose fine-grained details, or fail to complete occluded regions. In contrast, our method preserves the visible appearance and aligns with the pose condition. (C, D): OccThuman2.0 test examples under severe occlusions, where previous methods often produce implausible or pose-misaligned results. In contrast, our method synthesizes occluded regions that remain consistent with both the visible appearance and the pose condition.
Figure S11 : In-the-wild qualitative results. We show in-the-wild results of our method on a grayscale MSCOCO example (A) and an unseen-identity MPII example (B).
Figure S12 : Severe-occlusion qualitative results. We show intermediate outputs ( Icc , Miv′ ) and the final output ( Irc ) of our method under severe occlusions. (A): AHP real example (76% occlusion). (B): OccThuman2.0 example (53% occlusion). Each row uses a different random seed.
LPIPS* ↓
KID* ↓
MSE* ↓
PSNR ↑
Joint Err. ↓
w/o dil
49.86
6.54
4.59
25.63
23.73
w/ dil
49.47
6.12
4.37
25.86
23.33
Table S11 : Quantitative comparison of our method with and without dilation. Invisible mask dilation is used as a conservative margin for boundary uncertainty and has a marginal impact on metrics.
With the widespread adoption of autonomous vehicles and robotics, amodal completion, which reconstructs the occluded parts of people and objects in an image, has become increasingly crucial. Just as humans infer hidden regions based on prior experience and common sense, this task inherently requires physical knowledge about real-world entities. However, existing approaches either depend solely on the image generation ability of visual generative models, which lack such knowledge, or leverage it only during the segmentation stage, preventing it from explicitly guiding the completion process. To address this, we propose AmodalCG, a novel framework that harnesses the real-world knowledge of Multimodal Large Language Models (MLLMs) to guide amodal completion. Our framework first assesses the extent of occlusion to selectively invoke MLLM guidance only when the target object is heavily occluded. If guidance is required, the framework further incorporates MLLMs to reason about both the (1) extent and (2) content of the missing regions. Finally, a visual generative model integrates these guidance and iteratively refines imperfect completions that may arise from inaccurate MLLM guidance. Experimental results on various real-world images show impressive improvements compared to all existing works, suggesting MLLMs as a promising direction for addressing challenging amodal completion.
Personalized image completion aims to restore occluded regions in personal photos while preserving identity and appearance. Existing methods either rely on generic inpainting models that often fail to maintain identity consistency, or assume that suitable reference images are explicitly provided. In practice, suitable references are often not explicitly provided, requiring the system to search for identity-consistent images within personal photo collections. We present AlbumFill, a training-free framework that retrieves identity-consistent references from personal albums for personalized completion. Given an occluded image and a personal album, a vision-language model infers missing semantic cues to guide composed image retrieval, and the retrieved references are used by reference-based completion models. To facilitate this task, we introduce a dataset containing 54K human-centric samples with associated album images. Experiments across multiple baselines demonstrate the difficulty of personalized completion and highlight the importance of identity-consistent reference retrieval. Project Page: https://liagm.github.io/AlbumFill/
Generating realistic and structurally plausible human images into existing scenes remains a significant challenge for current generative models, which often produce artifacts like distorted limbs and unnatural poses. We attribute this systemic failure to an inability to perform explicit reasoning over human skeletal structure. To address this, we introduce SkeleGuide, a novel framework built upon explicit skeletal reasoning. Through joint training of its reasoning and rendering stages, SkeleGuide learns to produce an internal pose that acts as a strong structural prior, guiding the synthesis towards high structural integrity. For fine-grained user control, we introduce PoseInverter, a module that decodes this internal latent pose into an explicit and editable format. Extensive experiments demonstrate that SkeleGuide significantly outperforms both specialized and general-purpose models in generating high-fidelity, contextually-aware human images. Our work provides compelling evidence that explicitly modeling skeletal structure is a fundamental step towards robust and plausible human image synthesis.