2026 ยท Under submission
Troupe: Scaling Multi-Identity Generation to Crowds on a Frozen Few-Step Backbone
D. A. Nguyen, J. Chavan, D. Lee, et al. (UIUC, Cantina Labs)
Up to 20 people in one image, each from a single photo. Highest identity similarity on a new 2-to-20-person benchmark, closed models included, at about 37% of the strongest open baseline's FLOPs.
