Garment transfer that keeps the pose — and the fabric.
Diffusion-based virtual try-on pipeline: garment transfer onto a person photo with pose preservation and fabric detail — built for e-commerce visualization.
This work was built for a private engagement, so the code and full pipeline aren't public. What follows is the approach at a level I can share, with a recorded result.
A recorded run.
The pipeline, in stages.
Human parsing
Segment the person into semantically consistent regions so garment placement respects body parts and clothing boundaries.
Pose extraction
Estimate keypoints so the transferred garment follows the person's stance instead of the reference garment's pose.
Garment transfer
Warp the target garment onto the person with pose-conditioned alignment.
Diffusion refinement
A refinement pass restores fabric texture, folds, and lighting consistency that geometric warping flattens out.
Where the difficulty actually lived.
Not in the models — in the seams between them: keeping long-range garment geometry stable across poses, preserving fine patterns that parsing smears, and holding identity constant through the diffusion pass. Success was measured on e-commerce realities — sleeves, prints, and layered clothing — not on studio photos.