Add MuLaCover F16 and Q8 GGUF packages

.gitattributes

@@ -266,3 +266,5 @@ PocketTTS-Small-GGUF/english-2m/english-2m-f32.gguf filter=lfs diff=lfs merge=lf
 PocketTTS-Small-GGUF/english-2m-refl/english-2m-refl-f32.gguf filter=lfs diff=lfs merge=lfs -text
 PocketTTS-Small-GGUF/english-7m/english-7m-f32.gguf filter=lfs diff=lfs merge=lfs -text
 PocketTTS-Small-GGUF/english-7m-refl/english-7m-refl-f32.gguf filter=lfs diff=lfs merge=lfs -text
+MuLaCover-GGUF/mulacover-f16.gguf filter=lfs diff=lfs merge=lfs -text
+MuLaCover-GGUF/mulacover-q8_0.gguf filter=lfs diff=lfs merge=lfs -text

MuLaCover-GGUF/MODEL_LICENSE

@@ -0,0 +1,64 @@
+MuLaCover Model Weight and Output Terms
+
+Copyright (c) 2026 MuLa Labs. All rights reserved except as expressly granted
+below.
+
+1. Scope
+
+These terms apply to all MuLaCover checkpoint weights distributed by MuLa
+Labs, including the MuLaCover-oss model weights and any fine-tuned, converted,
+quantized, merged, or otherwise modified versions derived from those weights
+(collectively, the "Model Weights"). They do not replace the licenses that
+apply to MuLaCover source code or third-party components.
+
+2. Model Weight License
+
+The Model Weights are licensed under the Creative Commons
+Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0):
+
+https://creativecommons.org/licenses/by-nc/4.0/legalcode
+
+You may copy, share, and adapt the Model Weights only for noncommercial
+purposes and only in compliance with CC BY-NC 4.0, including its attribution
+requirements. Attribute the Model Weights as "MuLaCover by MuLa Labs" and
+include a link to the official MuLaCover repository:
+
+https://github.com/HeartMuLa/MuLaCover
+
+3. Generated Output Restriction
+
+In addition to the Model Weight License, and as a condition of accessing or
+using the Model Weights, you may use outputs generated with the Model Weights
+only for noncommercial purposes. "Outputs" include generated or transformed
+audio, music, MIDI, tokens, latent representations, and other material produced
+through use of the Model Weights.
+
+Without a separate written commercial license from MuLa Labs, you may not sell,
+license for a fee, monetize, commercially distribute, commercially publish, use
+in advertising, use in a paid product or service, or otherwise exploit Outputs
+for commercial advantage or monetary compensation.
+
+These output terms are separate terms governing use of Outputs and access to
+the Model Weights. They do not modify the text of CC BY-NC 4.0. They apply only
+to the extent MuLa Labs has rights or contractual authority to impose them and
+do not limit rights that cannot lawfully be restricted.
+
+4. Commercial Licensing
+
+Commercial use of the Model Weights or Outputs requires a separate written
+license from MuLa Labs. Contact MuLa Labs through the official repository for
+commercial licensing inquiries.
+
+5. Third-Party Rights and User Responsibility
+
+These terms grant no rights in third-party source recordings, compositions,
+lyrics, performances, datasets, trademarks, or other material. You are solely
+responsible for obtaining all permissions required for your inputs and Outputs
+and for complying with applicable law. Third-party components remain subject
+to their respective licenses and notices.
+
+6. Disclaimer
+
+TO THE EXTENT PERMITTED BY LAW, THE MODEL WEIGHTS AND OUTPUTS ARE PROVIDED
+"AS IS", WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND. MULA LABS WILL NOT BE
+LIABLE FOR CLAIMS, DAMAGES, OR OTHER LIABILITY ARISING FROM THEIR USE.

MuLaCover-GGUF/README.md

@@ -0,0 +1,83 @@
+---
+library_name: audio.cpp
+license: cc-by-nc-4.0
+base_model:
+- HeartMuLa/MuLaCover
+- Qwen/Qwen3-Embedding-0.6B
+- HeartMuLa/HeartCodec-oss-20260123
+tags:
+- audio.cpp
+- gguf
+- music
+- text-to-audio
+- audio-to-audio
+---
+
+# MuLaCover GGUF
+
+Follow or star [audio.cpp](https://github.com/0xShug0/audio.cpp) for the native inference implementation.
+
+MuLaCover generates cover songs from reference audio or melody/chord MIDI, with
+new lyrics and style conditioning. These packages contain the complete pipeline:
+MuLaCover, the Qwen3 style encoder, HeartCodec, YourMT3, and five ChordNet folds.
+
+The base checkpoint is pinned to
+[HeartMuLa/MuLaCover at bbbaef2b31835c2ef5172ff528dbe325230d46fb](https://huggingface.co/HeartMuLa/MuLaCover/tree/bbbaef2b31835c2ef5172ff528dbe325230d46fb).
+The reference implementation is
+[HeartMuLa/MuLaCover at e27fa7fa0da7c9817879cc6210cc5693f9327946](https://github.com/HeartMuLa/MuLaCover/tree/e27fa7fa0da7c9817879cc6210cc5693f9327946).
+
+> [!IMPORTANT]
+> Upstream restricts the model weights and generated outputs to noncommercial use.
+> Read [MODEL_LICENSE](MODEL_LICENSE). The Apache-2.0 code license does not grant
+> commercial rights to the weights or generated audio.
+
+## Packages
+
+| File | Storage |
+|---|---|
+| `mulacover-f16.gguf` | FP16 floating weights, FP32 vector parameters |
+| `mulacover-q8_0.gguf` | Q8 eligible matrix weights across components, FP16 convolutions and lookup tables, FP32 vector parameters |
+
+Both files embed the model contract, configuration, tokenizers and chord vocabulary.
+Use the `mulacover` family. Generation is offline; progress through audio frames
+is not a streaming waveform API.
+
+## Download
+
+```bash
+hf download audio-cpp/audio.cpp-gguf \
+  --include "MuLaCover-GGUF/*" --local-dir models
+```
+
+## Generate from MIDI
+
+```bash
+audiocpp_cli --task gen --family mulacover \
+  --model models/MuLaCover-GGUF/mulacover-q8_0.gguf --backend cuda \
+  --text '[Verse] A morning light across the sea. [Chorus] Keep a little hope for me.' \
+  --request-option 'tags=topic:[hope]; genre:[folk]; instrument:[piano]; mood:[warm]' \
+  --request-option melody_file=melody.mid \
+  --request-option chord_file=chord.mid \
+  --request-option seed=1234 --out cover.wav --log
+```
+
+Add `--request-option drum_file=drums.mid` for drum conditioning. MIDI timing
+uses musical ticks; its tempo is not a separate target-tempo control.
+
+## Generate from reference audio
+
+```bash
+audiocpp_cli --task gen --family mulacover \
+  --model models/MuLaCover-GGUF/mulacover-q8_0.gguf --backend cuda --audio reference.wav \
+  --text '[Verse] A morning light across the sea. [Chorus] Keep a little hope for me.' \
+  --request-option 'tags=topic:[hope]; genre:[folk]; instrument:[piano]; mood:[warm]' \
+  --request-option seed=1234 --out cover.wav --log
+```
+
+Tempo is estimated automatically. Set `--request-option bpm=120` to supply the
+reference tempo explicitly. Reference audio and MIDI inputs are mutually exclusive.
+Upstream selects the transcribed lead-singing track as melody; input without a
+detected melody is rejected. Reference-audio preprocessing requires SOXR support.
+
+`duration_sec` defaults to a 300-second maximum generation budget. The model can
+stop earlier at EOS. The prompt and frame budget must fit the 8,192-token context.

MuLaCover-GGUF/mulacover-f16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a73ce63ab2c4b8a9bd606d40ccf83f95ed68ecc8ac2c6210e45cf2dc56e3b57
+size 13380996416

MuLaCover-GGUF/mulacover-q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5daca50ce959c63dcfef3be1e3a7cd855eee753a91cab1a3475da1af781e4518
+size 8220378496