InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators

TL;DR

InstantFusion proposes a shared latent interface for connecting heterogeneous image generators, enabling seamless denoising transitions across models. This interface allows for flexible cross-model state translation and achieves competitive performance in various settings. It enables new applications such as acceleration and cross-architecture distillation.

Why It Matters

InstantFusion has practical significance as it enables the creation of more diverse and flexible image generation systems. It allows for the combination of different models and techniques, leading to improved performance and new applications. This can be particularly useful in fields such as computer vision and graphics.

Key Findings

• cross-model compatibility of latent representations
• shared latent space for denoising trajectories
• flexible cross-model state translation