thepatch/melodyflow-t24-30secs-GGUF

MelodyFlow T24 30secs GGUF

GGUF conversion of facebook/melodyflow-t24-30secs for audiocraft.cpp's native melodyflow-server, which runs terry, gary4local's MelodyFlow audio editing service.

The three files are a complete bundle: put them in one folder and point melodyflow-server --models-dir (or AC_MODELS_DIR) at it.

File	Size	What it is
melodyflow-dit-t24-1.0B-v1.0-F16.gguf	1,836 MiB	the flow-matching DiT, F16
melodyflow-vae-48khz-v1.0-F32.gguf	228 MiB	the 48 kHz stereo VAE, F32
t5-base-encoder-0.1B-v1.0-F16.gguf	210 MiB	T5-base text encoder for prompts, F16

$env:AC_DEVICE = "cuda"
melodyflow-server --port 8002 --models-dir C:\path\to\this\folder

F16 is what fits beside a 30-second edit's activations on an 8 GB GPU, and it's a legitimate tier rather than a compromise: against an F32 PyTorch reference it lands about as far away as the CUDA reference itself does. The measurements are in MELODYFLOW_DIT.md and MELODYFLOW_EDIT.md; the routes match the Python service gary4juce already talks to, per SERVICES.md.

Sources and licensing

• DiT and VAE: facebook/melodyflow-t24-30secs revision 77bcfce24371bf29a06152c72169.
• T5 encoder: google-t5/t5-base revision a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1, Apache 2.0.

MelodyFlow's weights are released by Meta under CC-BY-NC 4.0, and this conversion inherits that license: non-commercial use only. T5-base is Apache 2.0. Each GGUF records its own source and license in its general.* metadata.