AIPLAY STUDIO

Make a song.
Make it yours.

Generate the track, shape its score, mix it in the DAW and bring it to the screen. Music, covers, vectors, video, VFX and 3D in one free, open-source studio. Work by hand or let your own MCP assistant use the same project tools.

Start with Music only for native YuE2, without ComfyUI or Python. Choose Full Studio for the wider suite. Local generation needs no Studio account or credits; optional hosted services use your own key and may charge.

Download for Windows Shape your track What you can make Music-only start Choose your setup GitHub

One place for the song and the project around it. Install the engines you choose.

“Make It Yours”: the 31-second Studio tour. Music, editable scores, the DAW, Pictures and Production. Made locally through MCP with a YuE2 song, animated Studio screenshots and native VFX.

Keep creating after the first render

The first take is a starting point.

Compare performances, reshape a musical passage and give the important moment a visual accent. Each workflow has a UI, a saved receipt and typed MCP controls.

♩

Notes into a new take

Edit a melody in the DAW, review its YuE2 score and try two or three alternatives for selected bars. Keep the rest of the recording with 80 ms blends at the seams, and compare before keeping. Compatible saved Python YuE2 songs only; this is background generation, not instant audio inpainting.

DAW score and AI-take workflow

◧

Hear the difference

Group 2, 4 or 8 saved songs. Switch A–H at the same playback position with loudness matched for listening, then save your choice, favourite and receipts. Different seeds are queued for generation separately.

Take comparison

✦

One moment, two timelines

Place a Drop cue at a bar and beat. Preview a dry impact or 909 kick with a flash, then apply editable objects to the DAW and VFX composition. Check the explicit time offset; Undo removes the unchanged cue objects.

Shared cues

Make music on your own machine. Local, private, no credits.

Start with Music only

YuE2 music without ComfyUI or Python.

Choose Q4 or optional Q8

Q4_0 is the smaller default: about 2.93 GB of model files. Q8_0 is higher precision: about 4.53 GB. Both share an F16 VAE and four sidecars. Runtime downloads are separate: about 833 MB for NVIDIA, 58 MB for Vulkan or 24 MB for CPU. Install explicitly in Models; source and integrity checks travel with the manifest.

49.4 seconds of audio in 22.0 seconds

One short-song test on an RTX 4070 Ti SUPER with 16 GB VRAM, CoT full and 32 synthesis steps. This is one measured run, not a promised speed or long-song quality comparison.

Controls that match this engine

CoT full, melody or off; 32 synthesis steps by default, with 16 experimental. Import ABC notation and plan bars, BPM and nominal length in Music or through MCP. Planning uses no GPU; changing a score generates a new take, not seamless extension. Choose installed Q4/Q8 weights per take in Music or through MCP; no silent substitution. Nonempty lyrics required; reference audio and instrumental support depend on the route. See the guide for this native engine's current limits.

Sampled whole-GPU peak: 6,589 MiB, baseline 3,129 MiB. Other applications are included: this is not a process-memory peak or certification for 6 GB or 8 GB cards. YuE2 weights use CC BY-NC 4.0. Studio's output label follows the authors' statement that individuals may sell output and companies need a commercial licence; review the linked licence details before publishing. Q8 has not yet been GPU-benchmarked or listening-tested in Studio; higher precision is not a guaranteed quality upgrade. Read the benchmark and licence limits.

The optional full suite

A music video, made end to end on one desktop.

The song, the words, the singing, every frame, the cover — all of it generated on a single consumer graphics card, in this program. No cloud service was called, no credits were spent, and nothing was uploaded.

“Hex Appeal” — 3:28, forty-four scenes. The song, the singing, every shot and the edit came out of this program, on one desktop with a 16 GB graphics card. Nothing was uploaded and nothing was paid for.

“Measure Twice” — 1:35. Written, sung, illustrated, filmed and cut by AIPLAY Studio. 100% AI-generated, 100% locally: the track came out of the music model, the shots out of the video models, and the edit out of the timeline — on the same machine, offline.

Every frame of it is recorded. Each render on this page went through the engine door, which wrote the graph, the prompt verbatim, the seed, the model files and the hash of every reference before the card spent a second on it. That is what makes “100% AI-generated” a checkable statement here rather than a slogan.

“Interruption at Altitude” — 80s, fifteen shots, one cast. Every frame is MiniMax H3 at its native 1344×768, 20 steps. The three characters hold together across all fifteen because their descriptions are repeated verbatim in every prompt — consistency is a writing discipline here, not a feature. Point a phone at the last five seconds: the city resolves into a working QR code that opens this page. It is generated art, not an overlay, and it is verified to decode after H.264 under blur and downscale.

And the pieces, on their own

A song, words and all

“Glass and Neon” — 2:37, written and sung on a home GPU from one paragraph of direction.

A video clip

Five seconds at 1280×704, from a sentence.

The same picture, three ways

An audio reference re-rendered at three denoise settings — the dial that decides how much of the original survives.

The separate Python YuE2 integration

A song you can read before you hear it.

YuE2 plans a lead sheet — two voices and the chords, in plain ABC — before it sings a note. Studio keeps that sheet where you can read it, change a chord or the tempo, render again, and print it as sheet music. Twelve songs from it, each rendered on the same 16 GB card as the older full-suite examples on this page. These are not native GGUF benchmarks; editable score export belongs to the Python integration.

“Hold the Morning Back”

Big-room festival house · 2:45

“The Long Way Round”

Acoustic world-folk · 2:22

“My Own Machine”

Anthemic electronic pop · 1:47

“I Live Here Now”

Symphonic gothic rock · 2:24

“The Morning Is Now”

Gospel-adjacent R&B anthem · 2:37

Past three minutes on a 16 GB card.

The same card stopped at 2:48 the day before. The reason turned out to be one default in the model's own code — the synthesis prefill's attention ran over the whole song in a single block, so its memory grew with the square of the length — and Studio now passes a 512-token block instead. These seven rendered on that card the same night, from Create, with their sheet music beside them. The first one is the song that ran out of memory at 3:14 the hour before.

“Count the Windows”

Classic pop-soul · 3:14 · the plan that OOM'd at the default

“One More Morning”

Post-grunge alternative rock · 3:46

“For a Minute I Was Certain”

Uplifting trance · 4:14

“Motel Chandelier”

Slow cinematic torch ballad · 4:25 — the longest so far

“The Third Stair”

Solo acoustic singer-songwriter · 3:44

“The Lamp I Leave Lit”

Slow ethereal female vocal · 3:55

“Four Minutes of Weather”

Warm indie pop · 3:45 · typed into Create and pressed — lyrics, a style, nothing else

Every one of these is original: the lyrics were written for it, and the tune, the chords and the arrangement are the model's. YuE2's weights carry CC BY-NC 4.0. Studio keeps a conservative noncommercial / not-for-sale label; this is not a legal finding that every generated song is covered adapted material. The older Python memory figures do not establish a native GGUF hardware minimum.

The whole surface

What you can actually make

The full suite goes beyond native music-only. These additional tools may need ComfyUI, Python and their own model downloads.

One studio, song to video: the track you just wrote, and the timeline you cut it against.

♪

Songs with vocals

Generate from lyrics and a style, keep the track in your library and create its cover. Choose the engine per song; references, instrumentals and continuation depend on the selected route.

MiniMax Music 3 · YuE2 3B · ACE-Step 1.5, with LoRAs and covers

♩

A song with a score

Read and edit saved YuE2 ABC, print a lead sheet or move melody notes into the DAW. Expand the main player and choose Score for note highlighting and Follow scrolling, with timing adjustments for notation that drifts from the recording.

Saved ABC · following player · MIDI · DAW

▦

Covers, layers and vectors

Generate cover art, edit layers, masks, selections and text, or turn a raster into SVG. Review detail crops and clean selected geometry through the same UI and MCP workflow.

Vector workflow · Qwen · FLUX · other installed engines

▶

Video clips

Generate clips from text or pictures, then use supported continuation and control routes to develop the shot. Pick the engine and render settings for your machine.

MiniMax H3 · LTX 2.5 · WAN VACE

✂

Videos and episodes

Plan scenes, cast characters with reference sheets, storyboard, keep alternative takes and assemble to a timeline. Start a music video from a song, or carry a shared bible and chosen references between episode projects.

The Production screen · video and audiobook projects

☾

Batch and schedule

Collect different song ideas from Music, retain each one's lyrics, style and engine settings, and choose how many takes to make. Start now or schedule later while Studio runs and the computer stays awake.

Multiple ideas · separate seeds · one song at a time

▤

A real DAW

Piano roll, tempo and meter maps, instruments, recording, insert racks, sends and returns. Finish with EQ, dynamics and stereo tools, then inspect loudness, true peak and delivery checks.

Builtin synthesis · optional instruments · CPU rendering

◈

A compositor

Build compositions with layers, keyframes, effects, masks, mattes and nested comps. Preview the result and keep editing by hand or through MCP.

The VFX screen

◎

Video that moves with the song

Use analysed music timing to drive cuts, pulses and transitions. Compositor-only styles need no video model; motion and frame-repainting styles use their own optional models.

Reactive video · editable VFX compositions

⚭

Making something with friends

Exchange sealed project files and reviewed scene-render requests with trusted friends. Returned takes enter quarantine before you adopt them. The file-based lending path is built; a two-PC end-to-end trial remains to be done.

Collab and reviewed handoffs

◎

Community sessions

Find live listening rooms and upcoming events, or borrow a style pack for your next song. Community stays at the top of the sidebar, with a live-session indicator from the public feed.

Public sessions · events · reusable styles

◉

Camera control

Block a shot in Blender, then use its preview as control footage for the supported VACE route. Control quality depends on the scene and model.

Blender previz → WAN VACE

⬢

Pictures into 3D

Build a mesh from a picture and preview 3D avatars. Mesh generation needs its own models and hardware; the optional rigging runtime currently cannot run on Windows.

TripoSG · avatar preview · optional UniRig

✦

A local assistant

Use Simple mode, Chat and Enhance to develop lyrics, style and creative requests. Choose a local writing model or connect an API assistant in Settings.

Optional local model or your own provider key

◆

Or drive it from your own assistant

Connect an MCP-capable assistant to the same project workflows. Inspect readiness, create assets, edit scores and scenes, compare receipts and apply reviewed changes. UI playback, OS dialogs and some trust decisions remain human controls.

MCP workflow map · configuration in Studio's Agent screen

⚙

Plan, replay and train

Use optional native planning and saved-stage replay, prepare a training dataset, train adapters, transcribe audio to MIDI and preview processing. These advanced tools have separate runtime requirements; the current planner/trainer needs NVIDIA CUDA.

Native workbench

Own the tools and the workflow

Local first. Clear choices.

Choose where work runs

Local generation and editing run on your computer with no Studio account or per-render credits. Optional hosted engines and API assistants send their required inputs to the provider you choose. Paid services use your own key.

Choose the models

Studio ships no model weights. Models shows download sizes, hardware requirements, licences and known territorial restrictions before installation. Apache-2.0 covers Studio's code; model and output rights depend on the selected model.

Inspect the work

Read saved prompts, seeds, settings and provenance where recorded. Compare alternatives, preview edits and keep the version you choose. Older recordings may have incomplete receipts; matching a seed alone does not guarantee identical audio.

It also publishes its own negative results. Where a feature is weaker than it sounds, the documentation says so and shows the measurement — including the ones that did not go our way.

A look around

The program itself

The song creation screen — Write a song — style, lyrics, and a cover that matches.

The images screen — The image library and its editor.

The video screen — Clips, with the engine and size you choose.

The timeline screen — A timeline to cut picture against sound.

The models screen — Every model, its size, its licence and whether this machine can run it.

The batch planner in an earlier build — Plan a batch of ideas and takes.

Getting it running

Install

For native music, start with Node.js 20+ and Studio. No ComfyUI setup, Python environment or unrelated model download.

On Windows: download AIPLAY Studio Setup.exe and run it. It installs Studio, plus a private Node.js 20+ when this PC has none, with no admin prompt. Windows may say “Windows protected your PC” because the installer is not signed yet: click More info, then Run anyway.

Same public GitHub, two local launch modes. YuE2 music-only is part of AIPLAY Studio, not a separate repository. You download the app, then only the models you choose.

When Setup asks which build, choose Senzu's.

Music only

Native YuE2 GGUF songs on NVIDIA, AMD, Intel or CPU. No ComfyUI, Python or unrelated models. Start small; speed depends on your hardware.

Full Studio

The wider music, DAW, image, vector, video, VFX and 3D suite. Uses ComfyUI and Python for applicable tools. The launcher can find an existing ComfyUI or install a private one; optional packs have separate requirements.

Optional hosted services

Full Studio can use hosted MiniMax music with your fal.ai key. The separate Use Comfy API launch mode offers hosted pictures, clips, sound and 3D with your Comfy key; it has no Music or Production screen. Inputs go to the provider, and its usage charges apply.

Source code (for developers):

git clone https://github.com/Senzube4n/AIPLAY-Studio.git
cd AIPLAY-Studio
npm ci --omit=dev
npm run start:music

Or download and extract the source ZIP (for developers; it needs Node.js installed by hand). Either way, start AIPLAY Studio.exe and choose Music only. In Models, explicitly install YuE2 GGUF Q4 or Q8 and its native runtime, then enter a style and lyrics in Music and press Create. The installer shows the download total and notices; unavailable or unverified packages must not appear ready.

What your machine needs

Native music prerequisite	What is known
OS / GPU	Windows x64. Setup selects CUDA for NVIDIA, Vulkan for AMD/Intel, or CPU. A 16 GB RTX 4070 Ti SUPER was timed; native Vulkan/CPU performance and a minimum are not certified.
System components	Current graphics drivers. NVIDIA needs CUDA 13.3 compatibility and the Microsoft Visual C++ v14 x64 Redistributable. Vulkan/CPU packages carry their runtime DLLs.
Application	Node.js 20+ and Studio. No ComfyUI, Python or PyTorch for native music-only.
Disk	About 2.93 GB for Q4 or 4.53 GB for Q8 model files. Runtime: about 833 MB for NVIDIA, 58 MB for Vulkan or 24 MB for CPU. Allow room for extraction and output files.

The models it can fetch, and on what terms

Nothing is bundled and nothing downloads on its own. Every row below is generated from the catalogue the app itself reads, so the licence and the size on this page are the licence and the size the Models screen will show you. Choose one music engine; the full-suite capabilities below are not prerequisites for native music-only.

Capability	Download	Licence
Music engine — MiniMax Music 3	11.9 GB	MiniMax Music3 Community
Music engine — YuE2 GGUF Q4 / optional Q8 (experimental)	~2.9 GB	CC BY-NC 4.0 (weights) · Apache-2.0/MIT (native code) · NVIDIA CUDA runtime terms on NVIDIA only · ⚠ sellable by individuals
Cover — SheetSage2 song-to-score (ComfyUI)	1.4 GB	CC BY-NC 4.0 (weights) · ⚠ not for sale
Video clips — TaoMate 3-step LoRA (H3)	2.5 GB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Video clips — video-to-video control (H3 Fun ControlNet)	2.3 GB	MiniMax H3 Community Licence (a patch on H3's weights) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Video clips — Fast setting for H3 (TaoMate 3-step, 182 MB)	182 MB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Video clips — 4-step speed-up for H3 (1.96 GB)	2.0 GB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Video clips — 8-step speed-up for H3 (1.96 GB)	2.0 GB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
H3 conditioning bridge — BUNNY (action logic)	22.0 MB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
H3 conditioning bridge — Semantic Bridge v1	11.0 MB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Music engine — YuE2 3B for ComfyUI (int8)	4.0 GB	CC BY-NC 4.0 (weights) · ⚠ sellable by individuals
YuE2 instrumental planner LoRA (ComfyUI)	213 MB	CC BY-NC 4.0 (weights, derived from YuE2-3B) · ⚠ not for sale
YuE2 real-audio NAR LoRA (ComfyUI)	108 MB	CC BY-NC 4.0 (weights, derived from YuE2-3B) · ⚠ not for sale
YuE2 real-audio tokenizer (head + MERT-v2-FullSong)	2.7 GB	CC BY-NC 4.0 (head from YuE2-3B; MERT-v2-FullSong) · ⚠ not for sale
Music engine — ACE-Step 1.5 turbo (ComfyUI)	14.7 GB	MIT (ACE-Step's LICENSE; the ComfyUI repack's card tags apache-2.0 with no licence text)
Music engine — YuE2 3B	7.8 GB	CC BY-NC 4.0 (weights) · ⚠ sellable by individuals
Audio reference — MiniMax Music 3 DAV encoder	306 MB	MiniMax Music3 Community · + a pip package
Cover art — FLUX.2 klein 4B	12.5 GB	Apache-2.0
Fast covers: Supra2-IMG	1.7 GB	Apache-2.0
Chat — Qwen3 4B	8.0 GB	Apache-2.0
Stem separation — HTDemucs (fine-tuned)	~336 MB	MIT · pip, not a download
Video clips — MiniMax H3 (quantised)	42.9 GB	MiniMax H3 Community · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Video clips — FastH3 (8 steps, experimental)	42.1 GB	MiniMax H3 Community Licence (derived from H3) · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Video references — MiniMax H3 ref2va	22.9 GB	MiniMax H3 Community · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
Background removal — BiRefNet	444 MB	MIT
Images — Ideogram 4 (open 9B)	25.2 GB	Ideogram Non-Commercial Model Agreement · ⚠ terms unread
Narration — TTS voices (Kokoro + Qwen3-TTS)	~15.5 GB	Apache-2.0 (both engines) · pip, not a download
Sound effects — Stable Audio 3 Small SFX	3.5 GB	Stability AI Community License
Images — Z-Image Turbo (Apache-2.0)	14.6 GB	Apache-2.0
Images — Qwen Image 2.1 (research licence)	17.3 GB	Qwen Research License Agreement · ⚠ not for sale
Images — Fast draft for Qwen Image 2.1 (Viggle turbo LoRA)	680 MB	Qwen Research License Agreement · ⚠ not for sale
Images — Krea 2 Turbo (community licence)	19.0 GB	Krea 2 Community License Agreement
Images — Z-Image base (Apache-2.0)	14.6 GB	Apache-2.0
Images — Anima (non-commercial model, sellable pictures)	1.4 GB	CircleStone Labs Non-Commercial v1.2
Video clips — LTX 2.5 (quantised)	39.7 GB	LTX-2.x Community · ⚠ gated repo
Structural control — WAN 2.1 VACE 1.3B	11.3 GB	Apache-2.0
Pose extraction — DWPose (TorchScript)	353 MB	Split: Apache-2.0 + terms unread · ⚠ terms unread
Depth extraction — Depth Anything V2 Small	99.2 MB	Apache-2.0
Depth extraction — Depth Anything V2 Large (non-commercial)	1.3 GB	CC-BY-NC-4.0 · ⚠ not for sale
Motion module — AnimateDiff v3 (SD1.5)	1.8 GB	Apache-2.0
Source on the hits — SparseCtrl RGB (AnimateDiff v3)	2.0 GB	Apache-2.0
SD1.5 checkpoint — DreamShaper 8 (the Motion look's painter)	2.1 GB	CreativeML Open RAIL-M
ControlNet v1.1 — depth and line art (SD1.5, fp16)	1.4 GB	OpenRAIL
Picture references — IP-Adapter Plus (SD1.5)	98.2 MB	Apache-2.0
CLIP vision tower — ViT-H/14 (LAION-2B)	2.5 GB	MIT
3D mesh from a picture — TripoSG 1.5B	7.9 GB	MIT
Skeleton and skin for a mesh — UniRig	5.8 GB	MIT
Whisper: transcription and timed lyrics	~3.1 GB	MIT · pip, not a download
Smooth motion — RIFE 4.26	22.7 MB	MIT
Upscale — Real-ESRGAN 2x	67.1 MB	BSD-3-Clause
MiniMax H3 fl2va W6A8 (experimental)	16.0 GB	MiniMax H3 Community · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
MiniMax H3 ref2va W6A8 (experimental)	16.0 GB	MiniMax H3 Community · ⚠ excludes European Union, United Kingdom, Republic of Korea, United States of America
YuE2 style LoRA — TRBDR Broadside	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — TRBDR Lantern	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — TRBDR Hearth	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — TRBDR Ferryman	177 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — TRBDR Porch	177 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — GRVL Ember	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — GRVL Smoulder	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — GRVL Cinder	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — GRVL Wildfire	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — GRVL Thunder	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale
YuE2 style LoRA — GRVL Tempest	118 MB	CC BY-NC 4.0 (adapter weights) · ⚠ not for sale

For the video models

GPU / RAM	Requirements vary by model, build, resolution and route. Review the selected model's card and memory requirements in Models before installation.
Disk	2.5 GB (TaoMate 3-step LoRA) or 2.3 GB (video-to-video control) or 182 MB (Fast setting for H3) or 2.0 GB (4-step speed-up for H3) or 2.0 GB (8-step speed-up for H3) or 42.9 GB (MiniMax H3) or 42.1 GB (FastH3) or 39.7 GB (LTX 2.5)

AMD or Intel card? The native preset installs audio.cpp's Vulkan build there, and its CPU build on a machine without a card (much slower). The full suite's other capabilities have separate requirements; do not infer native support from another engine's memory table. Read the native guide.

Local first
Apache-2.0 code
UI + MCP
No Studio account
Ships no weights
Optional hosted services

If you want the detail

Read deeper

DAW score and AI takesReviewed musical edits, alternatives with seam blends and current route limits. Take comparisonLoudness-matched A–H listening, receipts and saved choices. Shared cuesOne musical position, a builtin accent, a VFX flash and reversible Apply. Vector workflowRaster import, SVG tracing, selected geometry cleanup and visual review. Native music workbenchOptional planning, replay, dataset training, MIDI and processing tools. MCP workflow mapTyped tools and the UI workflows they share. YuE2 music-onlyInstall, native controls, source pins, the short-song benchmark and its limits. The engine doorHow render attribution and the shared engine path work. ProvenanceWhat the ledger stores, embeds and proves. The compositorHow to get something on screen, in order. Compositor contractThe composition document, layer schemas and effect stack. Image contractWhere committed pixels are decided. Resolution for facesMeasured render size and face detail. Reference bleedA documented video-reference limitation. Engine trapsKnown engine differences and failure cases. CollabSealed bundles, scene lending and reviewed returns. SourceAll of it. Apache-2.0.