goldhub/Ruach_Studio_Models

Ruach Studio · Models

Website · GitHub · The guide · ☕ Buy me a Coffee Machine

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:

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

or only what you need:

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

folder	files	size	for	source	licence
models/	YuE2-3B-BF16.gguf	6.7 GB	the music model (24 GB GPUs)	Serveurperso/YuE2-GGUF, from m-a-p/YuE2-3B	CC BY-NC 4.0
models/	YuE2-3B-Q8_0 · Q6_K · Q5_K_M.gguf	3.6 · 2.8 · 2.5 GB	the same, quantized (12–16 GB GPUs)	as above	CC BY-NC 4.0
models/	YuE2-Vae-F32.gguf	0.5 GB	the sound decoder (default)	m-a-p/YuE2-Vae, converted	CC BY-NC 4.0
models/	YuE2-Vae-legacy-F32.gguf	0.5 GB	the first decoder	m-a-p/YuE2-Vae-legacy, converted	CC BY-NC 4.0
models/	YuE2-Vae-blend-F32.gguf	0.5 GB	a 0.666 merge of the two	Mothersuperior/YuE2-Vae-merge-0.666, converted	CC BY-NC 4.0
models/	SheetSage2-F32.gguf · SheetSage2-Q8_0.gguf	2.6 · 0.9 GB	audio → score (covers)	m-a-p/SheetSage2 + m-a-p/MERT-v2-FullSong, converted	CC BY-NC 4.0
sliders/	16 sliders + catalog.json	0.2 GB	voice and genre sliders	ntc-ai/yue2-particle-sliders, converted	CC BY-NC 4.0
loras/YuE2-instrumental-cot-full-loras/	ar_lora_inst_v3abc.safetensors + its README and scripts	0.3 GB	instrumental pieces	Mothersuperior/YuE2-instrumental-cot-full-loras	CC BY-NC 4.0
whisper/whisper-large-v3-ct2-float16/	CTranslate2 float16	2.9 GB	the lyrics check, karaoke timing, trim to the text	Systran/faster-whisper-large-v3 @ edaa852, from openai/whisper-large-v3	MIT
separation/demucs/	htdemucs_ft (4 models + yaml)	0.3 GB	drums, bass, other, vocals	facebookresearch/demucs, as fetched by python-audio-separator	MIT
universr/	pytorch_model.bin, config.yaml	0.2 GB	upscale: redraws the top of the spectrum	woongzip1/universr-audio, code woongzip1/UniverSR	CC BY 4.0
listener/Qwen2.5-Omni-7B-GGUF/	Q8_0 · Q4_K_M · mmproj-f16	8.1 · 4.7 · 2.6 GB	the 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-7B	Apache-2.0
listener/Qwen2.5-Omni-7B-GGUF/	Q6_K · Q5_K_M · Q3_K_M · UD-Q2_K_XL	6.3 · 5.4 · 3.8 · 3.2 GB	the same, more sizes	unsloth/Qwen2.5-Omni-7B-GGUF @ de13a22	Apache-2.0
listener/Qwen2.5-Omni-7B-GGUF/	mmproj-Q8_0	1.5 GB	the audio projector, smaller	mradermacher/Qwen2.5-Omni-7B-GGUF @ 658a1c7	Apache-2.0
artwork/Krea-2-Muse/	museByStableYogi_v35Q8Extended.gguf · …Q4Extended.gguf	14.6 · 8.3 GB	paints 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 steps	Muse by Stable Yogi by Stable Yogi, the author's own GGUF files, unchanged	Krea 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 GB	what Krea 2 needs beside its transformer	krea/Krea-2-Turbo, unchanged	Krea 2 Community License
artwork/CyberRealistic-XL-v10/	diffusers, fp16	6.5 GB	the 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 fp32	CreativeML 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): Q8_0 + mmproj-f16 11.3 GB, Q4_K_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

what	why	where to get it
BS-Roformer ep317 (vocal stems)	its licence is not stated by its author	fetched 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 authors	their repositories, through the studio's tools/download-loras.sh
the raw checkpoints (m-a-p/YuE2-3B and the rest)	only needed to convert again	their repositories, tools/download-checkpoints.sh
Qwen3-4B-Instruct-2507 (7.6 GB), the artwork's prompt writer	it is already where it belongs, under Apache-2.0	Qwen/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™.