Adapters as Data: The Latent Space of LoRAs

The paper investigates treating LoRA adapters as a data modality by training a transformer autoencoder to map adapter weights into a learned latent space. It finds that this latent representation preserves the same capabilities as raw weight space—such as sampling, editing, and inversion—while offering greater robustness to displacement, producing distorted yet valid faces instead of texture loss. Additionally, the authors extend the model to accommodate mixed-rank adapters, enabling a single shared latent to host heterogeneous adapter architectures.