thepatch/foundation-1-GGUF

Foundation-1 GGUF

GGUF conversions of RoyalCities/Foundation-1 for sa3.cpp's optional native stable-audio-tools runtime. Foundation-1 is a full finetune of Stable Audio Open 1.0 and uses the same T5-base and Oobleck architecture.

Each encoding is a complete three-file bundle:

Tier	Complete bundle	Guidance
F16	2,377 MiB	reference and runtime default
Q8_0	1,296 MiB	conservative high fidelity
Q5_K_M	922 MiB	recommended compact tier
Q4_K_M	832 MiB	smallest footprint

All quantized tensors passed the per-tensor 0.990 cosine gate. Matched CUDA renders measured envelope/log-magnitude cosine of 0.9967/0.9901 at Q8, 0.9904/0.9550 at Q5, and 0.9643/0.9284 at Q4. All four tiers passed listening tests.

Musical timing

Foundation-1 was trained for 4- or 8-bar clips at 100, 110, 120, 128, 130, 140, or 150 BPM. The prompt should include the selected bar count and BPM. sa3.cpp's Foundation profile calculates the exact sample crop, the whole-second seconds_total conditioning value, and the padded Oobleck latent canvas independently to match RoyalCities' inference.

Unsupported model BPMs are intentionally rejected by this profile. The native sat-server accepts host_bpm, maps it to the nearest trained BPM, and pitch-preserves the output at the host tempo. That service-level stretch does not change the GGML model primitives; see SAT_SERVER.md.

RoyalCities' UI profile uses DPM++ 3M SDE, sigma 0.01–100, 100 steps, and CFG 7. The gary4local fallback profile uses DPM++ 2M SDE, sigma 0.5–50, 100 steps, and CFG 7.

Build the opt-in SAT CLI, download one complete tier, and generate directly from the Foundation musical grid:

cmake -S . -B build-sat -DSA3_BUILD_SAT=ON -DSA3_CUDA=ON
cmake --build build-sat --config Release --target sat-generate
python tools/download_models.py --sat --sat-model foundation-1
$env:SA3_DEVICE="cuda"
sat-generate --model foundation-1 --randomize --bars 4 --bpm 128 --seed 42 `
  --out foundation.wav

--randomize uses the RoyalCities M1 structured prompt vocabulary by default, then prints the exact prompt, seed, key, musical geometry, and output path. Use --randomize-mode mix for the richer T1 variant, --family Synth to lock its anchor, or --foundation-profile gary to A/B the gary4local fallback sampler. Run sat-generate --help for the complete model-specific option list.

Any omitted bars, BPM, key, or family value is randomized; explicitly supplied values act as locks. Without --randomize, pass --prompt as a manual descriptor override.

Audio-to-audio

stable-audio-open-oobleck-encoder-v1.0-F16.gguf is the Oobleck encoder, which audio-to-audio uses to turn input audio into the latent that sampling starts from. Generation never loads it, so it is a separate file from the decoder bundle above and comes in one F16 tier for every bundle (149 MiB). Its 146 tensors are identical whether converted from Foundation-1, Stable Audio Open 1.0 or Stable Audio Open Small; the GGML encoder's latent means and scales match PyTorch's with cosine 1.0000. sat-server's POST /audio2audio loads it when it is beside the other files; see SAT_SERVER.md. The published bundles are otherwise unchanged.

Sources and licensing

• Foundation-1 source revision: d3160956fa13a8f861d3f608ed24075d44e98554.
• Stable Audio Open 1.0 base revision: f21265c1e2710b3bd2386596943f0007f55f802e.
• T5-base revision: a9723ea7f1b39c1eae772870f3b547bf6ef7e6c1.

The Stability AI Community License is included as LICENSE.md; RoyalCities' source license notice is retained as LICENSE_FOUNDATION.md, and T5-base's Apache 2.0 license is included as LICENSE_T5.md. See NOTICE for conversion details.

Powered by Stability AI.