thepatch/foundation-1.2-keybeds-GGUF

Foundation-1.2 Keybeds GGUF

GGUF conversions of RoyalCities' Foundation-1.2 Keybeds checkpoint for sa3.cpp's native stable-audio-tools runtime. Keybeds is a further finetune of Foundation-1 trained to keep one sound's identity consistent across pitch, so a text prompt can become a playable instrument. It shares Foundation-1's architecture, T5-base encoder, and Stable Audio Open Oobleck autoencoder.

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	compact tier
Q4_K_M	832 MiB	smallest footprint

All quantized DiT tensors passed the per-tensor 0.990 cosine gate (Q8 178/0, Q5 175/0, Q4 175/0). Matched CUDA renders of two 12-note keybed previews (Rhodes and sine prompts, seed 42, same SDE noise streams) were sliced into 24 notes per tier and compared against F16:

Tier	Max pitch error	Octave errors	Envelope / log-mag cosine vs F16
F16	2.5 cents	0	reference
Q8_0	2.5 cents	0	0.9998 / 0.9985
Q5_K_M	2.5 cents	0	0.9992 / 0.9951
Q4_K_M	3.0 cents	0	0.9980 / 0.9894

These paired metrics measure divergence from one F16 trajectory, not absolute audio quality.

Keybeds

The model renders chromatic runs: a chunk of up to six notes, 3.0 s each with 0.25 s gaps, in one conditioned window (seconds_total 20 for six notes). sa3.cpp's sat/keybed.h follows RoyalCities' keybed tab and exporter (RC-stable-audio-tools 43dcbb4b):

Keybed, Sequence, Timbre Profile, Grand Piano, Warm, Dry, Chromatic Chunk, Note Sequence, C4, C#4, D4, D#4, E4, F4

• one seed shared by every chunk; RC's keyboards are prompt labels C2-B5, C2-F6, or C2-B6;
• DPM++ 3M SDE, 80 steps, CFG 6, sigma 0.03-500, no loudness normalization;
• slices on the 3.25 s grid, conservative -60 dB tail trim, 120 ms terminal fade.

Pitch. In our measurements every note renders exactly one octave below its prompt label (piano, sine, and bass prompts; "A4" in a sine prompt is a 220 Hz tone). sa3.cpp maps samples by sounding pitch, so a C2-B5 prompt range becomes MIDI keys C1-B4.

cmake -S . -B build -DSA3_BUILD_SAT=ON -DSA3_CUDA=ON
cmake --build build --config Release --target sat-generate
python tools/download_models.py --sat --sat-model foundation-1.2-keybeds
sat-generate --model keybeds --prompt "Rhodes Piano, Warm, Soft" --seed 1234 --out-dir kit

This writes one WAV per sounding note plus kit.sfz. A full 48-note kit took 79.5 s at F16 on an RTX 5070 Laptop GPU. --keybed-preview C4:6 renders a single chunk.

Through libsa3's C ABI, create a context with variant = "foundation-1.2-keybeds" and send one Generate request per chunk (duration_seconds = 19.25, conditioning_seconds_total = 20, sampler = SA3_SAMPLER_DPMPP_3M_SDE_V1). The SA3Keybed VST3/CLAP instrument does this and plays the result from MIDI.

Sources and licensing

• Foundation-1.2 Keybeds source: RoyalCities/Foundation-1 revision 7b10fbbbc1be2f54cbc5540aab89ee383bc94e4a, file Foundation-1.2-Keybeds.safetensors.
• 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.