InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators
Published in arXiv preprint, 2026
Xuanlang Dai*, Jiazi Bu*, Yujie Zhou, Yuhong Liu, Beichen Zhang, Penghui Yang, Ziyu Liu, and Yuhang Zang
Paper | Project | Code | Model
InstantFusion connects the denoising trajectories of heterogeneous diffusion and flow image generators. Model-specific Latent AutoEncoders project intermediate states into a shared latent space, trained to align states from the same image at matched noise levels and to synchronize their local denoising velocities.
Once trained, denoising can transition across frozen generators without a dedicated mapping for every model pair, supporting acceleration, multi-preference composition, cross-architecture on-policy distillation, and cross-model composition.
Recommended citation: Xuanlang Dai*, Jiazi Bu*, Yujie Zhou, Yuhong Liu, Beichen Zhang, Penghui Yang, Ziyu Liu, and Yuhang Zang. (2026). "InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators." arXiv preprint arXiv:2610.13944.
Download Paper