UNITE-AUDIO: Joint Learning of Continuous Tokenization and Latent Flow Matching for Text-to-Audio Generation
Paper • 2609.28206 • Published • 1
UNITE-AUDIO is a text-to-audio generation framework that jointly learns continuous tokenization and latent flow matching.
Checkpoints
| Checkpoint | Parameters | Description |
|---|---|---|
Stage 3 Flow Model (stage3_flow.safetensors) |
117.7M | The shared text-conditioned latent flow model. |
Default Decoder (default_decoder.safetensors) |
45.5M | The decoder checkpoint used for the reported paper metrics. |
Spectral Decoder (spectral_decoder.safetensors) |
45.5M | An alternative decoder with more stable high-frequency detail. |
The default setting uses the Stage 3 Flow Model with the Spectral Decoder. Parameter counts refer to the released checkpoint modules.
Quick Start
The official inference code downloads the required checkpoints automatically on first use.
Method Overview
Joint training with noisy partial-context flow matching.