Chameleon: Style-Content Disentangled Framework for Cross-Domain Object Compositing
Chameleon is a two-stage training framework for cross-domain image compositing, allowing foreground objects to be stylized to match different backgrounds while keeping their identity intact. It uses a disentangled encoder trained with Joint Hard Contrastive Learning and a diffusion transformer with Spatio-Temporal Attention Gating to achieve high compositional plausibility and stylistic fidelity.