Team Ai
Modelpublic

goldhub/Ruach_Studio_Models

sourceHugging Faceotherupdated 9h agoView on Hugging Face
3likes262downloads
Model Card

Ruach Studio · Models

<p align="center"><img src="https://ruachstudio.igr.bible/assets/ruach-banner.png" alt="Ruach Studio · OSEM: Open Source, Engaged &amp; Musical" width="760"></p>

<p align="center"><a href="https://ruachstudio.igr.bible">Website</a> · <a href="https://github.com/igrbible/RuachStudio">GitHub</a> · <a href="https://github.com/igrbible/RuachStudio/blob/main/docs/GUIDE.md">The guide</a> · <a href="https://buy.stripe.com/cNi4gz39t2R77bT6ey67S02"><b>☕ Buy me a Coffee Machine</b></a></p>

Every weight Ruach Studio uses, in one repository, laid out exactly as the studio's own folders, so one command puts each file where the studio looks for it:

bash
cd Ruach_Studio                      # the root of the studio
hf download goldhub/Ruach_Studio_Models --local-dir .

or only what you need:

bash
hf download goldhub/Ruach_Studio_Models --local-dir . --include "models/YuE2-3B-Q8_0.gguf" "models/YuE2-Vae-F32.gguf"

Nothing here is ours. These are the works of the people and teams credited below, re-hosted unchanged (the GGUF files as converted by yue2.cpp's tools, byte for byte the same as their sources'), so that a studio install does not have to visit a dozen repositories. Every file keeps its own licence; most are non-commercial (CC BY-NC 4.0). Using this repository you accept each licence of the files you take. If you are an author and want a file of yours taken down, open a discussion here and it goes.

What is here

folderfilessizeforsourcelicence
models/YuE2-3B-BF16.gguf6.7 GBthe music model (24 GB GPUs)Serveurperso/YuE2-GGUF, from m-a-p/YuE2-3BCC BY-NC 4.0
models/YuE2-3B-Q8_0 · Q6_K · Q5_K_M.gguf3.6 · 2.8 · 2.5 GBthe same, quantized (12–16 GB GPUs)as aboveCC BY-NC 4.0
models/YuE2-Vae-F32.gguf0.5 GBthe sound decoder (default)m-a-p/YuE2-Vae, convertedCC BY-NC 4.0
models/YuE2-Vae-legacy-F32.gguf0.5 GBthe first decoderm-a-p/YuE2-Vae-legacy, convertedCC BY-NC 4.0
models/YuE2-Vae-blend-F32.gguf0.5 GBa 0.666 merge of the twoMothersuperior/YuE2-Vae-merge-0.666, convertedCC BY-NC 4.0
models/SheetSage2-F32.gguf · SheetSage2-Q8_0.gguf2.6 · 0.9 GBaudio → score (covers)m-a-p/SheetSage2 + m-a-p/MERT-v2-FullSong, convertedCC BY-NC 4.0
sliders/16 sliders + catalog.json0.2 GBvoice and genre slidersntc-ai/yue2-particle-sliders, convertedCC BY-NC 4.0
loras/YuE2-instrumental-cot-full-loras/ar_lora_inst_v3abc.safetensors + its README and scripts0.3 GBinstrumental piecesMothersuperior/YuE2-instrumental-cot-full-lorasCC BY-NC 4.0
whisper/whisper-large-v3-ct2-float16/CTranslate2 float162.9 GBthe lyrics check, karaoke timing, trim to the textSystran/faster-whisper-large-v3 @ edaa852, from openai/whisper-large-v3MIT
separation/demucs/htdemucs_ft (4 models + yaml)0.3 GBdrums, bass, other, vocalsfacebookresearch/demucs, as fetched by python-audio-separatorMIT
universr/pytorch_model.bin, config.yaml0.2 GBupscale: redraws the top of the spectrumwoongzip1/universr-audio, code woongzip1/UniverSRCC BY 4.0
listener/Qwen2.5-Omni-7B-GGUF/Q8_0 · Q4_K_M · mmproj-f168.1 · 4.7 · 2.6 GBthe style listener: drafts a track's style tags for training sets (llama.cpp)ggml-org/Qwen2.5-Omni-7B-GGUF @ 89b7854, from Qwen/Qwen2.5-Omni-7BApache-2.0
listener/Qwen2.5-Omni-7B-GGUF/Q6_K · Q5_K_M · Q3_K_M · UD-Q2_K_XL6.3 · 5.4 · 3.8 · 3.2 GBthe same, more sizesunsloth/Qwen2.5-Omni-7B-GGUF @ de13a22Apache-2.0
listener/Qwen2.5-Omni-7B-GGUF/mmproj-Q8_01.5 GBthe audio projector, smallermradermacher/Qwen2.5-Omni-7B-GGUF @ 658a1c7Apache-2.0
artwork/Krea-2-Muse/museByStableYogi_v35Q8Extended.gguf · …Q4Extended.gguf14.6 · 8.3 GBpaints a take's artwork (player, cards, the MP3's cover): Krea 2 Muse by Stable Yogi v3.5 Extended, a fine-tune of Krea 2 Turbo, 8 stepsMuse by Stable Yogi by Stable Yogi, the author's own GGUF files, unchangedKrea 2 Community License
artwork/Krea-2-Turbo/text_encoder/ (Qwen3-VL-4B), vae/ (Qwen-Image), tokenizer/, scheduler/, model_index.json, transformer/config.json (no transformer weights: the fine-tune above brings its own)9.4 GBwhat Krea 2 needs beside its transformerkrea/Krea-2-Turbo, unchangedKrea 2 Community License
artwork/CyberRealistic-XL-v10/diffusers, fp166.5 GBthe artwork's painter on cards under 16 GB (Krea 2 wants about 15 GB at Q4, 21 GB at Q8, its prompt writer included)CyberRealistic XL v10 by Cyberdelia (SDXL 1.0 finetune), converted from fp32CreativeML Open RAIL++-M

Any LLM quant goes with either projector (mmproj-f16 or mmproj-Q8_0): llama.cpp takes them as a pair. Measured on one 120 s excerpt (RTX 3090, peak VRAM): Q80 + mmproj-f16 11.3 GB, Q4K_M + mmproj-f16 8.4 GB, about 5 s a track; the full bf16 model through transformers 18.3 GB. Every size heard the same tags: a draft to correct, not a description to trust.

Not here, and why

whatwhywhere to get it
BS-Roformer ep317 (vocal stems)its licence is not stated by its authorfetched by audio-separator at first use, or heresy/fetch-heresy.sh
yue2-industrial-rock-lora, yue2-steps-from-hell (monsterovich), YuE2_Deathmetalv1_lora (pduncan)no licence given: all rights stay with the authorstheir repositories, through the studio's tools/download-loras.sh
the raw checkpoints (m-a-p/YuE2-3B and the rest)only needed to convert againtheir repositories, tools/download-checkpoints.sh
Qwen3-4B-Instruct-2507 (7.6 GB), the artwork's prompt writerit is already where it belongs, under Apache-2.0Qwen/Qwen3-4B-Instruct-2507, pinned, through heresy/fetch-heresy.sh --artwork

Licences in short

  • —CC BY-NC 4.0 — share and adapt, with credit, not for commercial purposes. This covers YuE2, its decoders, SheetSage2, MERT, the sliders and the LoRA here. Their licences speak of the weights, not of the music made with them (YuE2's licence file: the weights, under CC BY-NC 4.0).
  • —Krea 2 Community License — Krea 2 and Krea 2 Muse (artwork/Krea-2-Turbo/, artwork/Krea-2-Muse/). By taking these files you agree to it and to Krea's Acceptable Use Policy; its copy (LICENSE.pdf) and the NOTICE.txt it asks for lie beside them. In short: use, copy and share them (with that licence and notice); commercial use only while your yearly revenue stays under one million US dollars; whoever deploys them must filter what is generated (section 4.2).
  • —CreativeML Open RAIL++-M — CyberRealistic XL (from SDXL 1.0): free use; its use restrictions (Attachment A, in artwork/CyberRealistic-XL-v10/LICENSE.md) bind everyone who uses it.
  • —Apache-2.0 — Qwen3-4B-Instruct-2507.
  • —MIT — Whisper large-v3 as converted by SYSTRAN (from OpenAI's weights, themselves Apache-2.0).
  • —MIT — Demucs (Meta AI Research).
  • —CC BY 4.0 — UniverSR weights.
  • —Apache-2.0 — Qwen2.5-Omni-7B and its GGUF conversions (its LICENSE beside them).

The full texts and every file's source are in LICENSE.md.

Credits

  • —YuE2, its decoders, SheetSage2, MERT — the Multimodal Art Projection team (m-a-p).
  • —GGUF conversions — Serveurperso (yue2.cpp, YuE2-GGUF).
  • —The YuE2 Kit these files were gathered by — IronWolve (yue2-kit).
  • —VAE merge, instrumental LoRA — Mothersuperior.
  • —Particle sliders — ntc-ai.
  • —Whisper — OpenAI; CTranslate2 and faster-whisper — SYSTRAN.
  • —Demucs — Meta AI Research; python-audio-separator — nomadkaraoke.
  • —UniverSR — woongzip1.
  • —Qwen2.5-Omni-7B — the Qwen team (Alibaba); its GGUF — ggml-org, unsloth, mradermacher.
  • —Krea 2 — Krea (Krea 2 is licensed under the Krea 2 Community License Agreement; see https://krea.ai/krea-2-licensing); its text encoder Qwen3-VL-4B and its VAE from Qwen-Image — the Qwen team.
  • —Krea 2 Muse — Stable Yogi, who modified Krea 2 into it.
  • —CyberRealistic XL — Cyberdelia.

Gathered for Ruach Studio by Viktor Zhuromskyy (ЙирмиЙа́Ѓу) with Claude (Anthropic) · HERETICAL TANDEM™.