Model card — zoo's first music generation

README.md

@@ -0,0 +1,70 @@
+---
+license: other
+license_name: stability-ai-community
+license_link: https://huggingface.co/stabilityai/stable-audio-open-small/blob/main/LICENSE.md
+tags:
+- core-ai
+- apple
+- on-device
+- text-to-audio
+- music-generation
+- stable-audio
+- diffusion
+base_model:
+- stabilityai/stable-audio-open-small
+pipeline_tag: text-to-audio
+---
+
+# Stable Audio Open Small — Core AI (on-device music generation)
+
+**The model zoo's first MUSIC / AUDIO generation model for Apple Core AI.** Type a prompt, get ~11s
+of 44.1 kHz stereo audio — generated entirely **on-device** on Apple Silicon. A community port of
+[`stabilityai/stable-audio-open-small`](https://huggingface.co/stabilityai/stable-audio-open-small)
+(Stability AI + Arm) to Core AI.
+
+A latent **diffusion** text-to-audio model: a T5 text encoder conditions a DiT (diffusion transformer)
+that denoises a latent over **8 rectified-flow steps**, then an Oobleck VAE decodes the latent to a
+waveform. Distilled (ARC) for few-step generation, so it's fast.
+
+## What's in the bundle (`macos/`)
+
+Three Core AI `.aimodel` bundles + a tiny host sampler loop:
+
+| bundle | role | I/O |
+|---|---|---|
+| `sa_cond_fp16b` | T5-base encoder + number conditioner | `input_ids[1,64], attention_mask[1,64], seconds_norm[1] → cross_attn_cond[1,65,768], global_embed[1,768], cond_mask[1,65]` |
+| `sa_dit_fp16` | diffusion transformer (run 8×) | `x[1,64,256], t[1], cross_attn_cond, global_embed, cross_attn_cond_mask → v[1,64,256]` |
+| `sa_vae_fp16` | Oobleck VAE decoder | `latent[1,64,256] → audio[1,2,524288]` |
+
+**Host loop** (`StableAudioRunner`): tokenize (T5, `t5_tokenizer/`) → conditioner → start from Gaussian
+noise → 8-step rectified-flow euler `x = x + (t_next − t)·v` over the fixed schedule
+`[1.0, .9944, .9845, .9579, .8909, .7455, .5125, .2739] → 0` → VAE decode → 44.1 kHz stereo wav.
+No KV cache, no CFG (cfg_scale 1.0 — the model is ARC-distilled).
+
+## Performance (M4 Max, GPU)
+
+| metric | value |
+|---|---|
+| 8-step DiT | ~200 ms (25 ms/step) |
+| VAE decode | ~185 ms |
+| **total** | **~0.4 s for ~11.9 s of audio (~30× real-time)** |
+| size | fp16, ~1.0 GB (DiT 651M + cond 210M + VAE 149M) |
+
+Numerics: each bundle engine-gated vs the reference at cos ≥ 0.9999; full pipeline reproduces the
+reference audio exactly.
+
+## Roadmap
+
+- iPhone (h18p) build — bundles AOT-compile; device RTF pending
+- int8 (further size cut)
+- a music-generation tab in the zoo app
+
+## Credits & license
+
+A community **Core AI conversion** — all credit to **Stability AI** (and Arm) for
+[Stable Audio Open Small](https://huggingface.co/stabilityai/stable-audio-open-small); T5 text encoder
+by Google. This bundle is governed by the **[Stability AI Community License](https://huggingface.co/stabilityai/stable-audio-open-small/blob/main/LICENSE.md)**
+(free for non-commercial use and for commercial use under \$1M annual revenue; review the license
+before use). No retraining — conversion only.
+
+Part of the [Core AI model zoo](https://github.com/john-rocky/coreai-model-zoo).