Create a local song visually with the official ComfyUI YuE2 template and exact 3B int8 checkpoint; publisher Linux wheel and community audio.cpp are available as alternate routes.
The GUI uses a 3.69 GB int8 file, but its full-song peak memory is not published here. Compare the entire pipeline, not checkpoint size alone.
| Machine | YuE2 · 3B int8 ComfyUI | YuE2 · 3B official | YuE2 · audio.cpp GGUF + WebUI |
|---|---|---|---|
| NVIDIA GPU · 16 GB VRAM | Checkpoint storage fits, but full generation memory on 16 GB is not verified; test a short song before relying on it. | Below publisher 24 GB GPU starting point; use a larger GPU or rent one. | Community GUI route may fit with a quantized package; check main model plus VAE memory. |
| NVIDIA GPU · 24 GB VRAM + 24 GB free RAM | 24 GB NVIDIA is a conservative starting class based on the publisher BF16 route; int8 ComfyUI memory has not been independently measured. | Publisher-supported starting class for a first song; check disk space. | Capacity candidate; confirm selected GGUF precision and VAE fit. |
| AMD Radeon · 24 GB VRAM | No validated YuE2 ComfyUI Radeon path in the cited template; do not assume CUDA instructions apply. | No documented YuE2 ROCm route; use the Linux NVIDIA rental. | No validated Radeon/YuE2 backend test here; check the exact audio.cpp Vulkan or HIP build. |
| Intel Arc · 24 GB VRAM | No validated YuE2 ComfyUI Intel path in the cited template; check node/runtime support first. | No documented YuE2 XPU route; use the Linux NVIDIA rental. | No validated YuE2 Intel backend test here; check the exact audio.cpp build. |
| Apple Silicon Mac · 64 GB unified memory | This guide does not claim the int8 template works on Apple Silicon; use the separate audio.cpp package only if its exact build supports your Mac. | No publisher MPS route; use the Mac as a rental client. | Metal runtime exists, but YuE2 peak memory and performance on this Mac require a real test. |
| CPU only · 64 GB RAM | CPU generation may be impractical; no measured ComfyUI CPU result is claimed. | No practical publisher CPU recipe; rent a Linux NVIDIA GPU. | CPU runtime exists, but full song generation may be impractically slow. |
| RunPod · NVIDIA GPU · 24 GB VRAM | Choose a rental based on full pipeline memory, not the 3.69 GB file size; start with 24 GB NVIDIA and test a 30-second song. | Meets the publisher GPU class if the pod also has 24 GB free host RAM. | GUI can run on a rented Linux host if the server and WebUI remain private. |
CC BY-NC 4.0 limits commercial use. Do not infer a speed estimate from the 3B parameter count.
The easiest first song uses ComfyUI’s exact YuE2 text-to-music template: install one checkpoint, enter a style and lyrics, then play the FLAC. The publisher Linux wheel and community audio.cpp WebUI remain alternate packages.
Select the weights you want. The instructions, downloads and settings change with your choice.
Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.
The checkpoint is 3.69 GB on disk, but that is not a peak-VRAM measurement. Check the full model, audio generation and host-memory use for your song length; no verified minimum or speed is claimed for this int8 package.
Choose your computer first, then its GPU path. Follow the visual setup and import the linked workflow.
On Windows, install the ComfyUI desktop app or NVIDIA build, then update it so the YuE2 nodes and templates are available. Open the local app after its installer finishes; these steps do not require PowerShell.
ComfyUI download and installation ComfyUI update guideIn ComfyUI, open Workflow Templates and select “YuE2: Text to Music”; if your build does not show it yet, download the official JSON and drag it onto the canvas. This is the exact Comfy-Org template for style-and-lyrics song creation, not an unrelated audio workflow.
Download the official YuE2 template JSON Inspect the official template and instructionsDownload yue2_3b_int8_convrot.safetensors (3.69 GB) and place it in ComfyUI/models/checkpoints/. In the checkpoint selector, choose that exact file. If ComfyUI offers to download the missing checkpoint automatically, confirm the same filename and wait until it is installed before queuing a song. This int8 ComfyUI package is distinct from the publisher’s BF16 wheel.
Choose the YuE2 3B int8 checkpoint, enter a music Style and short Lyrics with [Verse] and [Chorus], keep Full ABC Planning enabled, set Max Duration to 30 seconds and a seed such as 0, then click Queue. Max Duration is a cap; the actual song can end earlier.
Template controls and prompt guidancePreview the audio in ComfyUI, then check ComfyUI/output/audio/YuE2*.flac. Play the FLAC and confirm that the lyrics, vocals, accompaniment and ending make a complete song. Save the JSON with the output so you can revise Style, Lyrics or seed later.
Official YuE2 output and workflowA FLAC in ComfyUI/output/audio/ plays with vocals and accompaniment. Save the template and seed with the audio; alternate packages save their output in the locations named in their own steps.
Comfy-Org provides the template and checkpoint; YouRunAI has not physically measured this route on each listed GPU. The original publisher Linux wheel is a separate BF16 package with its own 24 GB starting requirement.
Check the actual 24 GB BF16-capable GPU and 24 GB free host RAM, close other GPU tasks, then choose a larger rental if the full pipeline still fails.
Use the exact wheel listed on the publisher card, confirm Hugging Face access and free disk space, then retry the model and VAE downloads.
Check the Linux Terminal output for a failed generation stage; look under the working folder for song.flac and outputs/song for artifacts before editing style or lyrics.
Check that the server build enabled the native model manager and that the YuE2 GGUF package, VAE and sidecars are present. Compare the WebUI and YuE2 model documentation; use the publisher Linux wheel path if your build cannot manage the converted package.
Update ComfyUI, load the linked official JSON if Browse Templates has not yet received it, then confirm the exact checkpoint filename appears in its selector. Stable desktop updates can lag template support.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.