Transferability of AudioGAR across generative models and conditioning settings

Determine whether the AudioGAR decoder-adaptation strategy transfers to additional latent generative models, codec architectures, and conditioning settings, including video-to-audio and image-to-audio generation.

Background

AudioGAR adapts a codec decoder using intermediate latents produced by perturbing encoder latents and denoising them through a frozen latent diffusion or flow-matching model. The reported experiments evaluate the method primarily on AudioX and TangoMusic for text-to-audio and text-to-music generation, although the decoder train-generation mismatch is observed across a broader collection of models and datasets.

The unresolved issue is whether the same adaptation strategy remains effective beyond the evaluated systems, particularly with other latent generative models, codec architectures, and conditioning modalities such as video-to-audio and image-to-audio generation. Establishing this would clarify the generality of AudioGAR and the extent to which its gains depend on the specific experimental configurations studied.