A paper proposes personalizing pose-driven human video diffusion from a single photo by generating its own training data, so no multi-pose video of the subject is needed.
Method: A pose-conditioned generator makes pseudo-ground-truth frames at key poses, a reference-guided sampler keeps the most identity-consistent ones, and a per-subject low-rank adapter trains on static clips built from them.
Claim: The synthetic frames, though lower in fidelity, supervise better than real ones, because real frames carry motion that erodes the temporal prior.
Results: On TikTok and UBC Fashion, the adapted model improves on the base on every metric, and it matches or beats adaptation on five real frames while using none.
Availability: Code and weights are linked, and the paper is accepted to WACV 2027.
Sign in to comment.
TL;DR of the paper, generated automatically.
A paper proposes personalizing pose-driven human video diffusion from a single photo by generating its own training data, so no multi-pose video of the subject is needed.
One photo, no subject video. Go figure.