ECCV 2026

ParallaxPortrait Matting

Xin Cai1,3 · Jiawen Chen2 · Lars Jebe2 · Tianfan Xue1,3,4 · Zhoutong Zhang2†

1CUHK   2Adobe NextCam   3Shanghai AI Laboratory   4CPII under InnoHK

Paper · PDF Supplementary Explore results ↓
Predicted foreground color for black-hair portrait Black-hair portrait input
↔
Drag to reveal foreground color Input frame Our foreground color
Predicted foreground color for colorful long-hair portrait Colorful long-hair portrait input
↔
Drag to reveal foreground color Input frame Our foreground color
One tiny camera move

Two frames.
One clean cut.

Single-image matting is fundamentally ambiguous. We turn the slight viewpoint change already present in burst photography into an extra constraint—separating fine foreground structures from complex backgrounds without a green screen.

01

Capture the parallax

Two portrait images from nearby viewpoints expose different foreground and background motion.

02

Align each layer

We estimate foreground and background motion separately, revealing complementary evidence near occlusion boundaries.

03

Fuse with confidence

Reliable background alignment is fused directly; noisier foreground alignment enters through cross-attention.

Architecture

Asymmetric
by design.

Our two streams share architecture but play different roles. The main path learns a stable matte from pixel-aligned background evidence, while the auxiliary path softly corrects uncertain foreground correspondences in feature space.

Overview of the Parallax Portrait Matting framework
Framework overview Background alignment provides trustworthy pixel-level fusion. Foreground alignment provides a flexible cross-attention cue that remains robust around hair and non-rigid motion.
4.13SAD on PPM-100
vs. 5.53 best prior
2.28MSE on PPM-100
lower is better
0.55Foreground MSE
on PPM-100
2×Casually captured views
no special hardware
Demo

See the
details move.

Watch how a small baseline creates enough motion parallax to recover hair strands, semi-transparent boundaries, and clean foreground colors in challenging real-world scenes.

Interactive gallery

Pixel-peep
the results.

Pick any two outputs, then drag the split view to inspect them at full resolution. Compare our alpha and foreground predictions with strong academic and commercial baselines.

Select two images to compare
Why it works

Motion makes
ambiguity visible.

After layer-specific warping, foreground and background evidence becomes easy to interpret: one alignment holds the background still, while the other holds the subject still.

Layer-specific motion alignment visualization
Complementary aligned views Disoccluded background regions and shifted background texture provide strong cues exactly where a single frame is underconstrained.
Resources

Read.
Watch. Cite.

Explore the camera-ready paper, supplementary material, and complete qualitative gallery. If this work helps your research, please cite it.

BibTeX

@inproceedings{cai2026parallax,
  title={Parallax Portrait Matting},
  author={Cai, Xin and Chen, Jiawen and Jebe, Lars
    and Xue, Tianfan and Zhang, Zhoutong},
  booktitle={European Conference on Computer Vision},
  year={2026}
}