InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators
The paper claims separately trained image generators can hand off denoising states through a shared latent space.
InstantFusion uses model-specific latent autoencoders to translate intermediate states from different generators into one shared interface. The authors train that space by matching states from the same image at similar noise levels and aligning local denoising velocities. Once trained, it lets frozen generators switch mid-trajectory without building a separate mapping for every model pair. They report uses in acceleration, preference composition, on-policy distillation, and cross-model composition. HF Daily Papers' note
InstantFusion uses model-specific latent autoencoders to translate intermediate states from different generators into one shared interface. The authors train that space by matching states from the same image at similar noise levels and aligning local denoising velocities. Once trained, it lets frozen generators switch mid-trajectory without building a separate mapping for every model pair. They report uses in acceleration, preference composition, on-policy distillation, and cross-model composition. HF Daily Papers' note
score 4