| This Stability AI Model is licensed under the Stability AI Community License, |
| Copyright (c) Stability AI Ltd. All Rights Reserved. |
|
|
| This is a Derivative Work. The Stability AI Materials were modified as follows: |
|
|
| The model was converted from its original PyTorch format to Apple's Core AI |
| `.aimodel` format, using `torch.export` and `coreai-torch`. The network was split |
| into four separately-callable graphs (text conditioner, diffusion transformer, |
| latent decoder, audio encoder) so that the sampling loop runs on the host. |
|
|
| Two training-time behaviours were disabled, both inactive at inference: |
| - `mask_noise` in the autoencoder (a training augmentation) |
| - the softnorm bottleneck's decode-time dither |
|
|
| The bottleneck dither was measured before removal: output with it disabled differs |
| from stock by SNR 18.8 dB / cosine 0.9934, while two stock runs differ from each other |
| by SNR 21.1 dB / cosine 0.9961. Removing it therefore lands within the model's own |
| run-to-run variance, and makes generation deterministic. |
|
|
| No weights were retrained, fine-tuned, quantized or otherwise altered in value. |
| Graph weights are stored at float32, matching the source checkpoint. |
|
|
| Original model: stabilityai/stable-audio-3-medium |
| The T5Gemma text encoder is subject to the Gemma Terms of Use (see LICENSE_GEMMA.md). |
|
|
