Amodal perception enables humans to perceive entire objects even when parts are occluded, a remarkable cognitive skill that artificial intelligence struggles to replicate. Recent diffusion-based methods have extended amodal completion from images to videos, yet they lack auxiliary priors to guide the reconstruction of heavily occluded content and to maintain temporal consistency over long sequences. We present REVEAL (REconstructing VidEo Amodal content via Language and flow guidance), a unified diffusion-based framework that integrates complementary motion and semantic priors for video amodal completion. A user-provided text query serves a dual role: it drives open-vocabulary video segmentation to obtain visible masks, and it provides semantic guidance for texture reconstruction. For amodal segmentation, we introduce optical flow guidance: by warping visible masks from previous frames to the current frame, the flow-warped mask propagates visible content into occluded regions and approximates the object's current shape, providing a strong shape prior even under simultaneous deformation and occlusion. For texture completion, text guidance constrains the reconstruction of occluded appearance while providing a stable semantic reference that helps maintain visual coherence throughout the reconstruction. We also contribute LAVAT, a new benchmark featuring long video sequences paired with text descriptions, enabling evaluation of text-guided video amodal completion under heavy occlusion. Extensive experiments demonstrate that REVEAL achieves state-of-the-art performance on existing benchmarks while maintaining high-quality temporal consistency over extended sequences. The project page is available at: https://uark-aicv.github.io/REVEAL/.
Figures & tables
Figure 2: Video Amodal Mask Segmentation Stage. The model uses optical flow as a motion prior, warping visible masks from previous frames to propagate shape information into occluded regions.
Figure 3: Video Amodal Texture Completion Stage. The second stage conducts amodal texture completion with text guidance as semantic priors, leveraging pretrained diffusion models to reconstruct accurate content based on canonical object shapes and properties.
Kubric
TACO-OvO
LAVAT (S)
LAVAT (R)
Video length
Short
Short
Long
Long
⋅ Min #frames
24
14
91
41
⋅ Max #frames
24
77
1595
661
⋅ Mean #frames
24
29
257
151
# videos
690
2049
2,658
100
# total frames
16,560
58,695
684,037
15,139
Table 1: Comparison of statistics with the existing video amodal completion datasets. For LAVAT , we report values for the synthetic subset (S) and real subset (R). # denotes “number of”.
Dataset
Method
Texture Completion
Segmentation
Temporal Consistency
PSNR( ↑ )
SSIM( ↑ )
LPIPS( ↓ )
CLIP-T( ↑ )
mIoU( ↑ )
CLIP-F( ↑ )
FVD( ↓ )
Kubric
Pix2Video
14.876
0.704
0.325
-
27.45
0.939
1073.13
E2FGVI
14.406
0.688
0.227
-
54.05
0.967
732.30
pix2gestalt
20.096
0.826
0.129
-
71.95
0.911
405.89
PCNet
20.111
0.861
0.111
-
79.70
0.958
253.77
Diff-VAS
21.213
0.852
0.099
-
81.05
0.974
229.04
Table 2: Quantitative comparison with state-of-the-art methods. Bold indicates the best performance.
Method
CLIP-T( ↑ )
User Pref. ( ↑ )
pix2gestalt ( Ozguroglu et al., 2024 )
0.274
6%
Diff-VAS ( Chen et al., 2025 )
0.279
12%
TACO ( Lu et al., 2025 )
0.280
11%
REVEAL (Ours)
0.298
71%
Table 3: Evaluation on LAVAT -R (real occlusions). We report text-video alignment (CLIP-T) and human preference (%) from a user study.
Figure 4: Qualitative comparison between TACO ( Lu et al., 2025 ) , Diff-VAS ( Chen et al., 2025 ) , and our REVEAL on LAVAT -R benchmark.
Setting
<20%
20 – 40%
40 – 65%
>65%
w/o flow prior
91.88
87.97
85.75
87.79
w/ flow prior
92.93
90.72
88.38
89.13
Table 4: Flow prior ablation across occlusion difficulty on LAVAT -S. We report amodal mIoU ( ↑ ).
Figure 5: Ablation of flow prior. Amodal mask prediction with and without flow prior on LAVAT .
Figure 6: Illustration of the flow prior under simultaneous deformation and occlusion. The object deforms while being occluded. The visible mask at frame T (green) covers only the unoccluded portion. Warping the previous frame’s mask to T via optical flow (purple) recovers regions that are currently hidden, providing a shape prior that captures the object’s motion and deformation. Our predicted amodal mask (yellow) closely matches the ground truth (orange).
Motion
Shape change
Setting
Low
Medium
High
Low
Medium
High
w/o motion guidance
89.55
85.28
76.26
90.36
85.72
77.70
w/ motion guidance (Ours)
92.48
86.84
78.10
93.30
87.87
80.73
Table 5: Motion guidance ablation across motion and shape-change magnitudes on LAVAT -S. We report amodal mIoU ( ↑ ), grouping motion and shape change separately.
Flow injection
Added params
Latency (ms)
<20%
20 – 40%
40 – 65%
>65%
w/o flow prior
0
86.9
91.88
87.97
85.75
87.79
Latent concat
11.5K
87.2 (+0.3)
92.20
88.50
86.40
88.10
MOFA-Adapter
735.2M
116.4 (+29.5)
92.80
90.30
87.95
88.80
Ours
0.635M
90.2 (+3.3)
92.93
90.72
88.38
89.13
Table 6: Flow-prior injection design ablation on LAVAT -S. Added params is on top of the 397.66M-parameter Stage-1 trainable set; Latency is per 25-frame 128 × 256 fp16 step on a single A6000. mIoU ( ↑ ) is reported across occlusion difficulty.
<20%
20 – 40%
40 – 65%
>65%
Setting
SSIM ↑
LPIPS ↓
CLIP-T ↑
SSIM ↑
LPIPS ↓
CLIP-T ↑
SSIM ↑
LPIPS ↓
CLIP-T ↑
SSIM ↑
LPIPS ↓
CLIP-T ↑
w/o text guidance
0.982
0.026
0.281
0.945
0.087
0.282
0.945
0.087
0.279
0.942
0.090
0.269
w/ text guidance
0.987
0.021
0.286
0.960
0.067
0.285
0.959
0.071
0.284
0.957
0.072
0.273
Table 7: Text guidance ablation across occlusion difficulty on LAVAT -S. We report completion quality for our Stage-2 texture model with and without text guidance.
Figure 7: Qualitative ablation of text guidance. Texture completion with and without text guidance on LAVAT . We also compare with VACE ( Jiang et al., 2025 ) as an additional video inpainting baseline.
Figure 8: Text cross-attention visualization. Given the prompt “A man that is wearing a white shirt, a cap and a bag”, we visualize the cross-attention maps for individual tokens. Each token attends to its corresponding semantic region, demonstrating that text guidance provides spatially grounded priors.
<250 frames
250–500 frames
>500 frames
Text guidance
CLIP-F ↑
FVD ↓
CLIP-F ↑
FVD ↓
CLIP-F ↑
FVD ↓
w/o text guidance
0.942
244.50
0.943
348.80
0.928
392.70
w/ text guidance
0.952
225.10
0.945
303.12
0.940
342.55
Table 8: Temporal consistency of text guidance across sequence lengths on LAVAT -S. We report Stage-2 texture completion results with and without text guidance using CLIP-F ( ↑ ) and FVD ( ↓ ). The with-text results are also reported in the “w/ Temporal Gen.” row of .
Occlusion duration (frames)
w/o flow prior
w/ flow prior
Δ
1–15
87.39
89.89
+2.50
16–40
88.03
90.61
+2.58
41–80
89.73
92.29
+2.56
81+
90.77
93.07
+2.30
Table 9: Flow prior ablation across occlusion duration on LAVAT -S. Amodal mIoU ( ↑ ) over occluded frames, grouped by the number of consecutive frames the object has been occluded.
Visible mask
Amodal mask → Stage 2
mIoU ↑
PSNR ↑
SSIM ↑
LPIPS ↓
FVD ↓
GT
GT amodal (Stage-2 upper bound)
—
29.55
0.972
0.058
276.4
GT
Stage-1 predicted (reported setting)
90.29
27.47
0.963
0.073
310.90
SAM3 ( Carion et al., 2026 )
Stage-1 predicted (end-to-end from text)
88.1
26.4
0.955
0.084
338.0
Table 10: End-to-end evaluation from plain-text input on LAVAT -S. The first row upper-bounds Stage 2 with ground-truth amodal masks; the second row is our reported setting; the third row runs the full pipeline from the text prompt alone, with visible masks predicted by SAM3.