LAE
Latent Autoencoder (LAE) is a model-specific lightweight autoencoder assigned to each generator within the InstantFusion framework. It performs bidirectional mapping between a generator's intermediate denoising states and a shared latent space, serving as the core component that enables cross-model collaboration among heterogeneous generative architectures.
Timeline (last 90 days)
InstantFusion introduces a model-specific latent autoencoder (LAE) for each generator, projecting intermediate states into a shared latent space and decoding them back into the latent space format required by the target generator.
The star topology design of the LAE reduces the total number of mappings from the N×(N-1) of traditional pairwise mappings to 2N, avoiding the combinatorial explosion problem.
All Facts (2)
InstantFusion introduces a model-specific latent autoencoder (LAE) for each generator, projecting intermediate states into a shared latent space and decoding them back into the latent space format required by the target generator.
50%UnverifiedThe star topology design of the LAE reduces the total number of mappings from the N×(N-1) of traditional pairwise mappings to 2N, avoiding the combinatorial explosion problem.
50%