SAM-Audio Small (AudioGhost copy)

An unofficial copy of Meta's SAM-Audio Small weights, repackaged for AudioGhost so that its users can download them without a Hugging Face account. Not made, endorsed or supported by Meta. The audio codec it needs is in audioghostai/sam-audio-shared.

What is different from Meta's checkpoint

  • Only the parts text-prompted separation uses: the diffusion transformer and the small layers around it (anchor embedding, projections, video alignment). The vision encoder (used only for visual prompting) is left out, and the audio codec lives in the shared repo. The text encoder is the public t5-base, as in Meta's code.
  • Safetensors instead of a pickled checkpoint.pt.
  • model.fp16.safetensors: the same tensors cast to float16. This is AudioGhost's default, and it loads into exactly the values that Meta's float32 checkpoint is cast to when AudioGhost runs it in float16.
  • model.fp32.safetensors: Meta's float32 values, unchanged (for bf16 / fp32 runs and fine-tuning).
  • config.json: Meta's, unchanged.

Source: facebook/sam-audio-small@20b65f56888142eebe7c37448c6f6b3b32600e9b.

License

These weights are SAM Materials under Meta's SAM License, included here as LICENSE. By downloading or using them you agree to its terms, including the Trade Controls section. If you redistribute them, or anything derived from them, you must do so under the same license and include a copy of it.

Files

File Size SHA-256
config.json 2,251 bytes (2251 B) 50a67c841676043afdd3af6484ef6a1b47e300e36ee24eb3e07c31da8216f47b
model.fp32.safetensors 1,984,936,272 bytes (1.85 GB) 7f099bfbd9ddd8dcdc45e15c3d881fc3990a663212600f0370b05b934a34c951
model.fp16.safetensors 992,482,380 bytes (946.5 MB) 907db246611b8743639ea847ebcb857db901a53db35aacfb799aa694a7a840b9
Downloads last month
2
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for audioghostai/sam-audio-small

Finetuned
(4)
this model