Controlla: Learning Controllability via Graph-Constrained Latent Geometry

Controllable multimodal generation is usually posed as inference-time conditioning, which leaves unstructured how semantic attributes evolve and invites identity drift. Controlla instead treats controllability as a property of latent geometry: it learns identity and attribute factors from multimodal inputs and aligns them with graph priors through graph-constrained optimal transport, so attributes follow graph-consistent trajectories while reference identity is preserved. The paper also contributes AffectHuman-43K, a leakage-aware multimodal benchmark.

Paper Project page