Diffusion and flow-based models have become the dominant approaches to image generation, producing a wide family of generators whose architectures and internal representations differ substantially. Yet how these models' latent representations relate to one another — and whether their denoising dynamics can be aligned — has remained largely unexplored.
In a new paper titled InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators, researchers report an intriguing cross-model compatibility: independently trained generators appear to share a similar latent spatial organization, and passing intermediate states directly between unaligned models can still preserve coarse visual content.
Building on this observation, the authors propose InstantFusion, a lightweight shared latent interface that connects the denoising trajectories of heterogeneous generators. Model-specific Latent AutoEncoders (LAEs) project intermediate states from different generators into a shared latent space and decode them into the latent space of a target generator, enabling flexible cross-model state translation.
The shared space is shaped by two complementary objectives: aligning intermediate states derived from the same image at matched noise levels, and synchronizing their corresponding local denoising velocities. Once trained, InstantFusion allows denoising to transition seamlessly across frozen generators without requiring a dedicated mapping for every model pair.
The authors demonstrate the interface across several settings — acceleration, multi-preference composition, cross-architecture on-policy distillation, and cross-model composition — reporting competitive performance in each. The work is presented as an initial exploration of whether generative models can share a latent space.