audiohacking/uniflow-audio-gguf

UniFlow-Audio GGUF

GGUF packs for uniflow.cpp — C++17 GGML inference for UniFlow-Audio (text-to-audio / text-to-music). No Python at runtime.

Small / Base / Large = DiT capacity.
F16 / Q8_0 / Q4_0 on dit-*.gguf = DiT weight encoding.
T5 is per pack (not one shared F32 blob): small uses Q4_0, base and large use Q8_0. F16 T5 is not published (unsupported on the CPU encoder path).

Recommended default: Base + dit-Q8_0.gguf.

Packs

Directory	Upstream	T5	DiT files
uniflow-audio-v1.1-small/	Small	Q4_0 (~183 MB)	F16 / Q8_0 / Q4_0
uniflow-audio-v1.1-base/	Base	Q8_0 (~346 MB)	F16 / Q8_0 / Q4_0
uniflow-audio-v1.1-large/	Large	Q8_0 (~346 MB)	F16 / Q8_0 / Q4_0

Also in every pack: vae.gguf, instructions.gguf, spiece.model (~93 MB).

Quick start

git clone --recursive https://github.com/audiohacking/uniflow.cpp
cd uniflow.cpp
make metal

# requires: pip install -U "huggingface_hub[cli]"
./scripts/download_gguf.sh                 # base + dit-Q8_0 (+ T5 Q8_0)

./build-metal/uniflow-audio --model base --quant Q8_0 \
  --caption "a man is speaking while a dog barks" \
  --output output/out.wav \
  --duration 5 --steps 25 --cfg 5.0 --sway -1 --seed 42

Defaults match the UniFlow Space T2A demo: CFG 5.0, steps 25, sway -1, 24 kHz.

See the uniflow.cpp README for full CLI options.

License

Upstream UniFlow-Audio is MIT. Follow Flan-T5 terms for t5_encoder.gguf / spiece.model. Conversion scripts live in uniflow.cpp convert/.