YuE2 instrumental cot full LoRAs

YuE2 instrumental cot full LoRAs

LoRA add-on for the YuE2-3B music model that makes it write instrumental tracks following a section plan such as intro, verse and chorus. For musicians already running YuE2.

Made by Mothersuperior on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a LoRA add-on for YuE2-3B, an AI music model. It makes the model write instrumental music that follows a section plan, such as intro, verse and chorus. It targets the autoregressive part of YuE2, with rank 64 across all 28 layers. It was trained on about 2,700 instrumental tracks.

Who it suits: It suits musicians and tinkerers who already run YuE2 and want instrumental tracks with a planned structure.

Industry | Who uses it | How they use it | Example
Music | Producers | Draft instrumental tracks with a chosen section order | Planning intro, verse, chorus, bridge, chorus, outro
Games and video | Content creators | Generate background music with a set structure | Making a dark ambient cue for a scene
Research | Music AI researchers | Study planned section structure in generation | Comparing timed and untimed section tags

You can skip it if you want songs with sung lyrics, since the training and the prompt format are for instrumentals. The license also allows non-commercial use only.

Where to use it: It works with the YuE2 pipeline, either through the author's script or inside ComfyUI.

• The YuE2 generation pipeline, with chain-of-thought mode set to full
• ComfyUI, using the standard LoraLoader node in the CLIP slot
• Windows, macOS or Linux, wherever YuE2 runs

Use case | Who | What they get back
Planned instrumentals | Producers | Tracks that follow the sections you list
Timed sections | Editors | A structure steered by start and end times, used as a guide
Free structure | Experimenters | The model picks the sections and length itself

How to set it up: You need YuE2-3B and a way to run it. Choose the file that fits your tool. The card lists a float32 file, a bf16 file at half the size, and a ComfyUI-layout file. A reference script merges the LoRA into the base weights.

• YuE2-3B and its pipeline
• One of the three LoRA files
• Optionally the author's NAR LoRA and tokenizer, for more realistic sound
• Python, for the Hugging Face download tool

1. Download the LoRA files from the repository.
2. Pick the file for your tool, with the ComfyUI file for ComfyUI.
3. Set the chain-of-thought mode to full in your pipeline call.
4. Write your style tags and section plan.

pip install -U huggingface_hub
hf download Mothersuperior/YuE2-instrumental-cot-full-loras

How to use it: Run YuE2 with the LoRA applied and chain-of-thought on, so the model writes its own score before the music. Give a style tag string, then put the structure in the lyrics field with bare tag names, one per line. The model tends toward songs of three to five minutes.

[instrumental]

[intro]
[verse]
[chorus]
[bridge]
[chorus]
[outro]

[intro 0:00-0:15]
[verse 0:15-0:45]
[chorus 0:45-1:10]
[bridge 1:10-1:40]
[chorus 1:40-2:05]
[outro 2:05-2:30]

Prompt form | What it does
Bare tag | The model chooses structure and length
Untimed tags | You set the section order and the model sets timing
Timed tags | Each section also gets its own time range

Example: You write a style string for dark ambient music with piano and theremin, then list intro, verse and chorus sections. The model writes a score first, then an instrumental track that follows your section order.

COT=full STYLE_TEXT="dark ambient, water phone, theremin, piano, slow" \
python scripts/ar_generate.py ar_lora_inst_v3abc.safetensors none out_tag any_style_track plan.txt 21

For newcomers and experts: If you are new, this is an add-on that only works on top of YuE2. Start with the bare instrumental tag to hear what it does, then try a plan.

If you know the field, it is a rank 64 LoRA on the attention and MLP projections, trained on instrumentals paired with chord-annotated ABC scores, half of the time with the score present. Held-out loss fell from 5.33 to 4.51 without a score and from 3.64 to 2.92 with one.

Style tags and section plan
Model writes an ABC score first
Music tokens follow that score
Stock NAR decoder renders audio

Good to know: Times in the timed form are a guide rather than a promise. The model follows section order and proportions better than the exact end time.

• License: CC BY-NC 4.0, so non-commercial use only
• Cost to run: no API key, account or paid service is named, but you need hardware able to run YuE2-3B
• Data it reaches: the prompts and plans you write, plus files downloaded from Hugging Face
• Not for: songs with lyrics text, production notes in the lyrics field, or commercial use

Title: Held-out loss without a score, lower is better
Before: 5.33
After: 4.51

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 95
• License: cc-by-nc-4.0
• Based on: m-a-p/YuE2-3B
• Tags: audio, music, yue2, lora, instrumental

Links

• Model: https://huggingface.co/Mothersuperior/YuE2-instrumental-cot-full-loras
• Files: https://huggingface.co/Mothersuperior/YuE2-instrumental-cot-full-loras/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Mothersuperior/YuE2-instrumental-cot-full-loras

Then follow the usage in its model card. Nothing in this kit runs by itself.