Ruach Studio · Models v2

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

Ruach Studio is open music generation that runs on your own hardware: the model, the score, the adapters and every take stay yours, and nothing is rented by the token. Open models are now good enough to stand on their own, and what was missing is the studio around them: that is what we build, in the open. Music should be free to make and to share: whoever has no voice, no instrument and no studio still has the right to lay their thoughts and feelings onto sound.

⚠ Not on free cloud GPUs. Ruach Studio does not run on Google Colab, Kaggle or any free notebook GPU, so don't spend an evening on it there: it is a whole studio, not a notebook. It builds its C++/CUDA engine on the machine it lives on, keeps a server and a lab of Python services running beside it, and takes about 33 GB of disk (60 GB with the Artist). A free session's card, memory, wiped disk and hours-long limit cannot hold that. It wants your own Linux computer, or a server you rent and keep, with an NVIDIA card: 24 GB for everything, 16 GB with the smaller model copies.

Everything Ruach Studio runs on, in one place and laid out as its own folders: the music model and its decoders, the transcriber, the stems' and the sound upscaler's models, Whisper, the style listener, the studio's own adapters, the LoRA Trainer's base, and the Artist's painters with the adapters their authors let be shared. 1662 files, 501 GB. A copy of the studio's own disk, file for file: what lies here is what it paints, sings and trains with, and the studio fetches all of it from here alone.

Nothing here is ours but the conversions and the studio's own adapters. Every other file is its authors' work, unchanged, under its authors' licence, credited below and line by line in MANIFEST.tsv. Not affiliated with M-A-P, Comfy Org, Mothersuperior, HaileyStorm, atomtanstudio, storagejuju, OpenAI, Systran, Meta, the Qwen team, Unsloth, lightx2v, Falconsai, prithivMLmods, Civitai or the authors named below. If you are an author and want a file of yours taken down, open a discussion here and it goes.

Take what you need

The studio's own install (./fetch-models.sh) takes what your card can run, at a pinned revision, checking each file. By hand, from the studio's root:

hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "models/YuE2-3B-Q8_0.gguf" "models/YuE2-Vae-F32.gguf" "models/SheetSage2-Q8_0.gguf" "models/MERT-v2-FullSong-Q8_0.gguf"   # to sing: 5.3 GB
hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "models/YuE2-Vae-*"                                                                   # the other decoders
hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "models/SheetSage2-Refined-F16.gguf" "models/SheetSage2-Refined/*"                 # the second transcriber (beside SheetSage2): 0.16 GB
hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "whisper/*" "separation/*" "universr/*"                                               # Whisper, stems, the sound's upscaler
hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "artwork/upscalers/*" "licenses/upscalers/*"                                    # the Artist's upscalers
hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "artwork/Qwen-Image-2512/gguf/*Q4_K_M*" "artwork/Qwen-Image-2512/lightning/*" "artwork/Qwen-Image-2512/*.json" "artwork/Qwen-Image-2512/scheduler/*" "artwork/Qwen-Image-2512/tokenizer/*" "artwork/Qwen-Image-2512/text_encoder/*" "artwork/Qwen-Image-2512/vae/*" "artwork/Qwen-Image-2512/transformer/*"   # the Artist on 16 GB
hf download goldhub/Ruach_Studio_Models_v2 --local-dir . --include "checkpoints/comfy/checkpoints/yue2_3b_bf16.safetensors" "checkpoints/MERT-v2-FullSong/*" "checkpoints/yue2-mothersuperior-realaudio-tokenizer-v4/*"   # the LoRA Trainer's base: 10.5 GB

Known by its content, not its name

SHA256SUMS holds each file's SHA-256 (as sha256sum -c SHA256SUMS reads it, and as Hugging Face knows the file). MANIFEST.tsv adds, for each .safetensors, the SHA-256 of its tensors alone (their names and data, its header's metadata aside), with each file's origin and licence. A file renamed after download, or an adapter whose header was rewritten, is the same file or the same adapter all the same: the studio looks for what it needs by these sums, not by the name it was published under.

What is here

folder	size	what for	from	licence
`models/`	22.1 GB	YuE2-3B (BF16, Q8_0, Q6_K, Q5_K_M), its decoders (Standard, Legacy, the 0.666 Blend) and the score's transcriber: SheetSage2's head with MERT-v2-FullSong beside it (F32, Q8_0 each), and Mothersuperior's realaudio tokenizer head (F32, Q8_0)	m-a-p/YuE2-3B and its siblings as GGUF (Serveurperso/YuE2-GGUF); Legacy and Blend converted by the studio from m-a-p/YuE2-Vae-legacy and Mothersuperior/YuE2-Vae-merge-0.666	CC BY-NC 4.0
`whisper/`	3.1 GB	Whisper large-v3 (CTranslate2, float16): the lyrics heard, the takes checked	Systran/faster-whisper-large-v3	MIT
`separation/demucs/`	0.34 GB	the stems (htdemucs_ft)	facebookresearch/demucs	MIT
`universr/`	0.23 GB	the sound's upscaler (the top of a take redrawn)	woongzip1/universr-audio	CC BY 4.0
`checkpoints/comfy/`, `checkpoints/MERT-v2-FullSong/`, `checkpoints/yue2-mothersuperior-realaudio-tokenizer-v4/`	14.5 GB	the LoRA Trainer's base: YuE2-3B whole in one file (bf16, or int8 for a smaller card), MERT-v2-FullSong (the set's latent cache hears with it) and the realaudio tokenizer head	Comfy-Org/YuE2, m-a-p/MERT-v2-FullSong (its `LICENSE` and notices beside it), Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4 (its card beside it)	CC BY-NC 4.0
`checkpoints/Qwen2.5-Omni-7B-GGUF/`	35.7 GB	the style listener (hears a training set and writes each song's tags) and the Artist's critic: six quants, two projectors	Qwen2.5-Omni-7B as GGUF: ggml-org, unsloth in `unsloth/`, mradermacher in `mradermacher/`	Apache-2.0 (Qwen's `LICENSE`, beside the GGUFs; ggml-org's card names it qwen-research, its link leads to the same Apache-2.0)
`checkpoints/Qwopus3.5-9B-GGUF/`	12.1 GB	the Writer's own language model, through llama.cpp's llama-server, text only: quality (the AutoRound 4 bits moved into Q4_0 bit for bit, what was kept in BF16 at Q8_0; 6.7 GB, a 12 GB card) and compact (the same 4 bits byte for byte, the embeddings Q4_K, the output Q6_K, the linear attention Q5_K; 5.5 GB, an 8 GB card)	Jackrong/Qwopus3.5-9B-v3, a fine-tune of Qwen/Qwen3.5-9B; quantized for Ruach Studio with AutoRound (W4A16, symmetric, group 32), the vision tower and the MTP head left out	Apache-2.0 (the `LICENSE` beside the GGUFs)
`loras/`	1.9 GB	the Kit's library of adapters, the ones their authors license (below); the studio's own: duduk and shofar, six voices by their kind (contralto, baritenor, bass)	the authors below; goldhub/Ruach_Studio_LoRAs	CC BY-NC 4.0
`artwork/Qwen-Image-2512/`	129.5 GB	the Artist's main painter: its parts, the GGUFs (Q4_K_M for 16 GB, Q6_K for 24 GB), Lightning (8 steps), and four fp8 single files for 24 GB: its own BASE and the fine-tunes IntoRealism, JibMix, Real	Qwen/Qwen-Image-2512, unsloth/Qwen-Image-2512-GGUF, lightx2v/Qwen-Image-2512-Lightning; the fine-tunes from Civitai by enzino, J1B and wikeeyang	Apache-2.0; the fine-tunes their authors' (share and sell)
`artwork/Qwen-Image-2.1/`	33.1 GB	the painter that letters and edits a finished picture	Qwen/Qwen-Image-2.1, unchanged, with its `LICENSE` and the `NOTICE` it asks for	Qwen Research Licence: research and evaluation only
`artwork/Qwen3-4B-Instruct-2507/`	8.1 GB	the covers' prompt writer	Qwen/Qwen3-4B-Instruct-2507	Apache-2.0
`artwork/nsfw-filter/`	0.34 GB	the covers' content filter	Falconsai/nsfw_image_detection	Apache-2.0
`artwork/upscalers/`	2.83 GB	the Artist's upscalers (a picture made 2 or 4 times as large, in passes): the 35 whose authors let them be shared	their authors on OpenModelDB and Civitai, below	CC BY 4.0, CC BY-NC-SA 4.0, WTFPL, BSD-3-Clause, Apache-2.0, CC0
`artwork/loras/`	7.5 GB	the Artist's adapters their authors let be shared: for 2512 Film Photography, Photostyle by Sean Archer (SFW), Sketch Smudge, Lightning in 4 steps (bf16, fp32) and 8 (fp32); for 2.1 Fantasy ArtStyle, Frank Frazetta, GPT-Image-2 Anime, Lenovo UltraReal	Civitai: bluejaysierra, FASCIUM, aimann, tarn59, makimakimakima12138, Danrisi; prithivMLmods/Qwen-Image-Sketch-Smudge; Lightning by lightx2v	each its author's: share and sell (Civitai); Sketch Smudge and Lightning Apache-2.0
`models/muscriptor/`	3.57 GB	the Refiner's sound to MIDI (MuScriptor small, medium, large; f16 GGUF), its `LICENSE` and `NOTICE.md` beside it	DamRsn/muscriptor-gguf, of MuScriptor by Mirelo and Kyutai	CC BY-NC 4.0
`checkpoints/YuE2-Vae/`, `checkpoints/YuE2-Vae-legacy/`, `checkpoints/YuE2-Vae-merge-0.666/`	1.59 GB	the three decoders as their authors publish them (safetensors), the sources of the GGUFs in `models/`, with their cards and licences	m-a-p/YuE2-Vae, m-a-p/YuE2-Vae-legacy, Mothersuperior/YuE2-Vae-merge-0.666	CC BY-NC 4.0
`checkpoints/MERT-v2-30s/`	2.53 GB	MERT-v2 for 30-second clips, kept beside MERT-v2-FullSong as m-a-p publishes it, its `LICENSE` and notices beside it	m-a-p/MERT-v2-30s	CC BY-NC 4.0
`checkpoints/SheetSage2/`	0.23 GB	the transcriber as m-a-p publishes it (safetensors and its code), the source of `models/SheetSage2-*.gguf`	m-a-p/SheetSage2	CC BY-NC 4.0
`models/SheetSage2-Refined-F16.gguf`, `models/SheetSage2-Refined/`	0.16 GB	the second transcriber: anime-song's SheetSage2-Refined (odd meters and meter changes, slash and tension chords, keys with modes), an overlay the engine reads with `models/SheetSage2-*.gguf` and its MERT beside it; its `LICENSE` and `NOTICE.md` in the folder	anime-song/SheetSage2-Refined, as GGUF by scragnog (scragnog/SheetSage2-Refined-GGUF)	CC BY-NC 4.0
`checkpoints/yue2-mothersuperior-realaudio-tokenizer-v4/` (the rest)	2.70 GB	the realaudio tokenizer's other heads and NAR adapters (v4, v5, v8, v9; fp32, bf16, ComfyUI), beside the head the LoRA Trainer reads	Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4	CC BY-NC 4.0
`checkpoints/YuE2-hum-to-song/`	1.60 GB	a hummed melody into a song: the prosody adapter and its regularizer pack, its card and scripts (not yet a room's button)	Mothersuperior/YuE2-hum-to-song	CC BY-NC 4.0
`checkpoints/audiobox-aesthetics/`	0.42 GB	Meta's aesthetics scorer (content enjoyment, usefulness, complexity, quality), kept for the Studio's probe measures	facebook/audiobox-aesthetics	CC BY 4.0 (its text in `licenses/upscalers/CC-BY-4.0.txt`)
`models/eagle3/`	0.15 GB	an EAGLE-3 draft head for the semantic phase of YuE2-3B (speculative decoding, not yet in the engine), its card beside it	engival/YuE2-3B-EAGLE3-draft-GGUF	CC BY-NC 4.0
`models/reader/resemble-enhance/`	0.71 GB	the Narrator's TTS Voice Upscaler: Resemble Enhance's enhancer (denoise and enhance a speaking voice)	ResembleAI/resemble-enhance	MIT
`checkpoints/particle-sliders/`	2.36 GB	the genre and voice sliders (rank-8 distilled v1 and v2, the particle weights), their card, loader and notes	ntc-ai/yue2-particle-sliders	CC BY-NC 4.0; the loader's code MIT (its `LICENSE`, `LOADER_LICENSE`)
`checkpoints/YuE2-3B/`	7.30 GB	YuE2-3B as m-a-p publishes it (safetensors, its code, card, licences and notices): the engine's GGUFs and the Trainer's single file are made from it	m-a-p/YuE2-3B	CC BY-NC 4.0
`checkpoints/Qwen2.5-Omni-7B-AWQ/`	12.72 GB	the style listener at 4 bits (AWQ), its `LICENSE` beside it	Qwen/Qwen2.5-Omni-7B-AWQ	Apache-2.0
`checkpoints/Qwen2.5-Omni-7B/`	22.38 GB	the style listener whole, at bf16 (the listener's `bf16` choice), its `LICENSE` beside it	Qwen/Qwen2.5-Omni-7B	Apache-2.0
`models/seedvr2/`	69.38 GB	the Videographer's and the Artist's upscaler: SeedVR2 3B and 7B (7B sharp the Studio's own) in fp16, fp8 and GGUF, and its VAE	ByteDance-Seed/SeedVR2-3B and -7B, as numz/SeedVR2_comfyUI, AInVFX/SeedVR2_comfyUI and cmeka/SeedVR2-GGUF pack it	Apache-2.0
`models/reader/drama-3-gguf/`	98.22 GB	the Narrator's voice: Drama 3 as GGUF for audio.cpp (Q8_0, experts Q6_K, experts Q4_K), its `LICENSE`, `NOTICE` and card beside it. Built with Fish Audio.	fishaudio/drama-3, converted by Ruach Studio (not an official Fish Audio release)	Fish Audio Community License 1.0 (its `LICENSE`, `NOTICE`)
`models/reader/dicts/uk/`	0.16 GB	the Narrator's Ukrainian stresses: 2.85 million word forms under their keys, the homographs kept with both stresses (named, never guessed), the names apart; its `NOTICE` names every link	lang-uk/ukrainian-word-stress (its stress dictionary, generated by Oleksiy Syvokon), from «Словники України» of the Ukrainian Lingua-Information Fund (ULIF, NAS of Ukraine); indexed by Ruach Studio	MIT (its `NOTICE`)

The fine-tunes and the Artist's adapters are here because their authors let them be shared. On Civitai each of these allows sharing and selling the model itself and its merges; they are credited all the same. Many others the Artist paints with allow running and renting but not sharing the file (Outpaint, Turbo 4 steps, High Resolution, More Real, Cinematic Photography, Flux Pro Style, the FinalCut fine-tune of 2.1 and more): they stay with their authors on Civitai; taken from there into artwork/loras/ or artwork/Qwen-Image-2.1/checkpoints/, the Artist offers them.

Not here yet, and why (11.11.2026). MiniMax H3 (the Videographer): its licence lets it be passed on only inside its own territory (not the US, the EU, the UK or South Korea), which a public repository cannot keep, so the studio fetches it from Comfy-Org and FastVideo. The Narrator's Russian stresses (A. A. Zaliznyak's accented paradigm): the rights to Zaliznyak's dictionary are Anna A. Zaliznyak's, and the paradigm the Narrator reads at home has no licence to pass on, so a Russian stress dictionary comes from its user (the Studio's docs/NARRATOR-DICTIONARIES.md says how). The stems' RoFormer and VR models: their licences are being checked. The Artist's newer Civitai adapters and fine-tunes: each one's sharing terms are being checked. The studio's own fine-tunes (RS-FW, RS-TW), its VAE-RS x3 and every voice are not in this repository by design.

The Artist's upscalers

Unchanged, each under its author's licence (texts in licenses/upscalers/). Left out: 8x NMKD-Faces and 8x NMKD-Superscale (no licence named) and GFPGAN 1.4 (parts under NVIDIA's non-commercial licence); the studio fetches none of them for you.

upscaler	for	author	licence
`1x_SkinContrast-SuperUltraCompact.pth`	Not an upscaler: the same size, more contrast in skin (backgrounds change a little). A pass of its own in a chain.	Brigadeiro	CC0-1.0
`4x-UltraSharpV2.safetensors`	UltraSharp V2 again: the same file as 4x_UltraSharp_v2. Slow. Non-commercial.	Kim2091	CC-BY-NC-SA-4.0
`4x_ClearRealityV1.pth`	Realistic pictures, best on faces, hair and nature; a soft, natural look with few artifacts, for fairly clean input (not grainy). Quick. Non-commercial.	Kim2091	CC-BY-NC-SA-4.0
`4x_ClearRealityV1_Soft.pth`	ClearReality's softer sister: kinder to rendered and game pictures, weaker on real photos. Quick. Non-commercial.	Kim2091	CC-BY-NC-SA-4.0
`4x_FaceUpDAT.pth`	Faces in good-quality photos. Slow.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_FaceUpSharpDAT.pth`	Faces in good-quality photos, sharper. Slow.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_NMKD-Siax_200k.pth`	An all-rounder for clean or lightly compressed pictures (JPEG quality 75 and up). Quick.	NMKD	WTFPL
`4x_NMKD-Superscale-SP_178000_G.pth`	Realistic photos with noise and compression. Quick.	NMKD	WTFPL
`4x_NMKD-UltraYandere_300k.pth`	2D art and anime, JPEG and dithering mended. Quick.	NMKD	WTFPL
`4x_NMKD-YandereNeoXL_200k.pth`	Anime and cartoons. Quick.	NMKD	WTFPL
`4x_Nomos2_hq_dat2.pth`	Clean photos (not compressed, not blurred): the most faithful detail on good input. Slow.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8kDAT.pth`	Photos, also JPEG-compressed or a little blurred: detailed and natural. The studio's default; slow.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8kHAT-L_otf.pth`	Degraded photos, faces too: a large model trained on blur, noise and compression. Slow, and the heaviest on the card.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8kSCHAT-L.pth`	Photos with JPEG compression and blur: a large model for realistic detail. Slow and heavy.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8k_atd_jpg.pth`	Photos with JPEG compression; keeps the grain and the noise. The slowest here, and in full precision only.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8k_span_otf_medium.pth`	Photos, quick: mends blur, noise and compression in the middle measure.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8k_span_otf_strong.pth`	Photos, quick: mends strongly (smoother). The first pass of Viktor's ComfyUI chain.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_Nomos8k_span_otf_weak.pth`	Photos, quick: mends little, keeps the most of the picture's own texture.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_RealWebPhoto_v4_dat2.pth`	Photos taken from the web: JPEG or WebP compression, blur and noise mended as it upscales. Slow.	Helaman (Philip Hofmann)	CC-BY-4.0
`4x_RealisticRescaler_100000_G.pth`	Small realistic textures crushed by JPEG or BC1, and outdoor photos. Quick.	Mutin Choler	WTFPL
`4x_UltraMix_Balanced.pth`	Natural pictures with sharpness and detail in balance. Quick. Non-commercial. The last pass of Viktor's chain.	Kim2091	CC-BY-NC-SA-4.0
`4x_UltraMix_Restore.pth`	UltraMix leaning to repair: damaged and compressed input. Quick. Non-commercial.	Kim2091	CC-BY-NC-SA-4.0
`4x_UltraMix_Smooth.pth`	UltraMix, smoother: soft surfaces, little sharpening. Quick. Non-commercial.	Kim2091	CC-BY-NC-SA-4.0
`4x_UniversalUpscalerV2-Neutral_115000_swaG.pth`	Realistic pictures, neither soft nor sharp. Quick.	Mutin Choler	WTFPL
`4x_UniversalUpscalerV2-Sharp_101000_G.pth`	Realistic pictures, sharper. Quick.	Mutin Choler	WTFPL
`4x_UniversalUpscalerV2-Sharper_103000_G.pth`	Realistic pictures, the sharpest of the three. Quick.	Mutin Choler	WTFPL
`4x_foolhardy_Remacri.pth`	General pictures with more detail and less smoothing: skin and fine texture do not turn mushy. Quick. Non-commercial. The third pass of Viktor's chain.	Foolhardy	CC-BY-NC-SA-4.0
`4x_foolhardy_Remacri_ExtraSmoother.pth`	Remacri, smoother: fewer edges and less grain. Quick. Non-commercial.	Foolhardy	CC-BY-NC-SA-4.0
`OmniSR_X4_DF2K.pth`	A research model trained on clean pictures: for clean input only. Small, yet slow. A free licence.	Hang Wang	Apache-2.0
`RealESR-General-WDN-x4v3.pth`	The tiny general model with denoising. The quickest here. A free licence.	Xintao Wang	BSD-3-Clause
`RealESR-General-x4v3.pth`	A tiny model for everyday pictures: the quickest here. A free licence.	Xintao Wang	BSD-3-Clause
`RealESRGAN_x2plus.pth`	Real-ESRGAN at 2×. Quick; a free licence.	Xintao Wang	BSD-3-Clause
`RealESRGAN_x4plus.pth`	The classic for real-world photos: compression, noise and blur mended. Quick; a free licence.	Xintao Wang	BSD-3-Clause
`RealESRGAN_x4plus_anime_6B.pth`	Anime pictures, a small model. Quick; a free licence.	Xintao Wang	BSD-3-Clause
`nmkdSiaxCX_200k.safetensors`	NMKD Siax again, as safetensors: the same model as 4x_NMKD-Siax_200k. Quick.	NMKD	WTFPL

The Kit's library of adapters

The adapters the YuE2 Kit offers, unchanged, at the revisions it pins, laid out as its own downloader lays them out (loras/<folder>/, the small side files the studio reads beside each):

adapter	what it brings	folder	author	licence
Instrumental	no vocals, planned sections	`loras/YuE2-instrumental-cot-full-loras/`	Mothersuperior	CC BY-NC 4.0
sv-billie	hushed bedroom pop	`loras/sv-billie-yue2-lora/`	HaileyStorm	CC BY-NC 4.0, its `LICENSE` beside it
DreamPop v2	hazy dream pop	`loras/sv-dreampop/`	atomtanstudio	CC BY-NC 4.0
J-pop	bright modern J-pop sound	`loras/yue2-jpop-t4-lora/`	storagejuju	CC BY-NC 4.0
Real audio v4, v9	real-audio sound, the older and the newest	`loras/yue2-mothersuperior-realaudio-tokenizer-v4/`	Mothersuperior	CC BY-NC 4.0
Industrial rock · Epic trailer · Death metal	riffs and mechanical tone · orchestra and choir · melodic death metal	not here	monsterovich (1, 2) · pduncan (3)	none stated: asked

What is not here, and why. Three adapters of the Kit's library carry no licence (Industrial rock, Epic trailer, Death metal): the studio fetches them from their authors' repositories, and they are not copied here until their authors say so (asked, linked above). The vocal stem's BS-RoFormer comes through audio-separator, and Qwen2.5-Omni-7B whole (the listener at bf16) from Qwen. The studio's private adapters (a person's own voice, a person's likeness) and its datasets are never published.

Licences

The strictest terms bind what you take: most of this repository is non-commercial. Every file keeps its source's licence, named in MANIFEST.tsv and in LICENSE.md; THIRD_PARTY_NOTICES.md groups them by origin, and licenses/ holds the full texts of the code the decoders name (stable-audio-tools and SnakeBeta/BigVGAN, MIT). MERT-v2-FullSong keeps its LICENSE (CC BY-NC 4.0, its full text) and THIRD_PARTY_NOTICES.md beside its weights. Qwen-Image 2.1 is used under the Qwen Research Licence: research and evaluation only, its notice kept. Using YuE2 in research, please cite YuE, as M-A-P asks.

Credits

YuE2, its decoders and SheetSage2: the Multimodal Art Projection team (m-a-p), as GGUF by Serveurperso. SheetSage2-Refined: anime-song, as GGUF by scragnog, whose HOT-Step CPP the engine's port follows. YuE2-3B in one file for ComfyUI: Comfy Org. MERT-v2-FullSong: m-a-p. The 0.666 decoder merge, the instrumental adapter, the realaudio tokenizer head and its NAR adapters: Mothersuperior. The Kit's library: HaileyStorm (sv-billie), atomtanstudio (DreamPop v2), storagejuju (J-pop), gathered by IronWolve's YuE2 Kit. Whisper: OpenAI, in CTranslate2 by Systran. Demucs: Meta AI. UniverSR: woongzip1. Qwen2.5-Omni, Qwen-Image 2512 and 2.1, Qwen3-4B: the Qwen team (Alibaba), the GGUFs by ggml-org, unsloth and mradermacher. Lightning: lightx2v. The content filter: Falconsai. The fine-tunes: enzino (IntoRealism Qwen), J1B (Jib Mix Qwen), wikeeyang (Real-Qwen-Image); 2512's own fp8 as Civitai publishes it. The adapters: bluejaysierra (Film Photography), FASCIUM (Photostyle by Sean Archer), aimann (Fantasy ArtStyle), tarn59 (Frank Frazetta Style), makimakimakima12138 (GPT-Image-2 Anime), Danrisi (Lenovo UltraReal), prithivMLmods (Qwen-Image Sketch Smudge). The conversions: convert-extras.py of the YuE2 Kit by IronWolve (yue2-kit), through the converter of yue2.cpp by Serveurperso.

Awesome YUE2: our curated list of everything the YuE2 community has made.

Gathered for Ruach Studio by Viktor Zhuromskyy (ЙирмиЙа́Ѓу) with Claude (Anthropic) · HERETICAL TANDEM™. The card before 14 October 2026 is kept as README.OLD.md.