| Model | Source | License |
|---|---|---|
| Whisper large/tiny | ggerganov/whisper.cpp | MIT |
| Kim Vocal 2 | Politrees/UVR_resources | MIT |
| RVC HuBERT (ContentVec) | lj1995/VoiceConversionWebUI | MIT |
| RMVPE pitch model | lj1995/VoiceConversionWebUI | MIT |
| wav2vec2-base-960h | Xenova/wav2vec2-base-960h | Apache 2.0 |
| Moondream2 vision | Xenova/moondream2 | Apache 2.0 |
| Piper voices | rhasspy/piper-voices | MIT |
| RobustVideoMatting (mobilenetv3) | PeterL1n/RobustVideoMatting | GPL-3.0 |
| YuNet (face detection) | OpenCV Zoo | MIT |
| Face Landmarker v2 + blendshapes | MediaPipe (tflite→onnx via tf2onnx) | Apache 2.0 |
| RVC pitch bases | generated from lj1995 pretrained_v2 f0G emb_pitch tensors | MIT (derivative) |
build/models/ — flat names for the
top-level models (vision/decoder_model_merged_q4.onnx → moondream-decoder.onnx,
vision_encoder_q4.onnx → moondream-encoder.onnx, embed_tokens_q4.onnx →
embed_tokens.onnx, tokenizer.json → moondream-tokenizer.json), face/ and
piper/ keep their subdirectories. hubert.onnx is exported from upstream
hubert_base.pt by the app's export-hubert tool (the .pt itself is not shipped
here — it is not used at runtime).