Upscale Audio (FLowHigh)

How it fits

Typical backends

Caveats

• Hard constraintGPU is effectively required; both the flow-matching estimator and the vocoder run on a CUDA device.
• I/O contractThe output is fixed at 48 kHz mono regardless of input rate or channel layout — downstream expectations must match. An empty input frame is passed through unchanged.
• I/O contractThe model reconstructs the missing high band for inputs in roughly the 8-24 kHz range; the input low band is copied through unchanged, so strong noise or codec artefacts in the input survive into the output.
• Parameter interactiontimestep is the number of flow-matching solver steps; the model is built for single-step inference, so the default of one is usually best and raising it costs latency for little gain.
• Hard constraintThe model slot must carry a FLowHigh flow-matching checkpoint plus a BigVGAN vocoder (its JSON config and matching weights) in one bundle; the worker loads both from the bound files at startup and never pulls from the network at run time.