YuE2 Music Generation Test: 4 Local Models Compared

👁 119 reads

YuE2 is an open model from the m-a-p team that turns lyrics and a style description into a complete song with vocals. It works natively in ComfyUI from version 0.41. I gave it the same Polish lyrics as three other local models (ACE-Step 1.5, MiniMax Music 3 and DiffRhythm 2), ran three attempts per model, then processed the recordings with Whisper. YuE2 sings clearly in Polish, with an average of 91.2% of the words recognized, but ACE-Step 1.5 edged it out with 95.6%.

Read the full guide to the Polish-language test winner: ACE-Step 1.5 running locally.

What is YuE2?

YuE2 is the second version of the YuE model, released on September 9, 2026. It has about 3 billion parameters. You give it lyrics marked with section labels ([Verse], [Chorus], [Bridge]) and a style description in English; it returns a finished song with vocals, instruments and structure. It can also create covers and edit sheet music in ABC mode, but I haven’t tested those features yet.

The original GitHub version requires about 24 GB of VRAM and Linux. In ComfyUI 0.41, it runs natively on Windows, with a ready-made template and weights repackaged by Comfy-Org (bf16 7.8 GB or int8 4 GB).

My test: 4 models, the same Polish lyrics

I wrote a short ballad with two verses and two choruses, 83 words in total, including Polish diacritics. The opening line, translated from the Polish test lyrics, is: “It’s raining over the Vistula again; trams glide into the gray mist…” I used the same style for all four models: “Polish pop ballad, warm female vocals, piano and soft synth pads, slow tempo”. Each model got three attempts with different seeds (11, 22 and 33). I ran the recordings through Whisper large-v3 with Polish selected, then calculated what percentage of the words from the lyrics appeared in the transcript in the same order. The test itself used Polish lyrics; the translated excerpt above is not a new test prompt.

This score measures intelligibility, not musical quality. Whisper sometimes picks up a word a person can’t understand, and sometimes misses one a person can. It is still useful for comparing models on the same lyrics.

What the scores show

• YuE2 had the best single attempt (98.8%, or 82 of 83 words), but also the most variation. In the attempt with seed 22, it repeated “[translation from the Polish original: By the Vistula]” (“By the Vistula”) twice and added phrases of its own. In the attempt with seed 33, it skipped several lines.
• ACE-Step 1.5 was the most consistent. It was the only one of the four models with a language selector in its settings, and I set it to “pl”. You set the song length in advance (120 seconds).
• MiniMax Music 3 sings clearly, but has two issues. It changes the order of the lyrics (two attempts started with the second verse, whose opening line translates as “Someone’s playing old jazz in Praga”) and cuts songs short: 73, 111 and 92 seconds with a 150-second limit.
• DiffRhythm 2 supports phonetics only for Chinese and English. It turns Polish lyrics into gibberish; one garbled phrase sounded roughly like “[corrupted lettering in the original Polish test]” in the Polish recording. Also, this model dates from October 2025, so it is not a new release from this year.

Listen: Polish lyrics, seed 11

English lyrics: YuE2 gets every word

The second set of lyrics was in English, with an ’80s synth-pop style and male vocals. I ran one attempt per model.

Generation time and VRAM use

The times cover everything from submitting the job to getting the file, including model loading when ComfyUI switched between models. Generation itself is faster, as ACE-Step shows: it took 8 seconds when the model was already in memory. MiniMax Music 3 uses the entire 24 GB graphics card and slows down as a result.

How to run YuE2 in ComfyUI

1. Update ComfyUI to version 0.41 or later.
2. From the Comfy-Org/YuE2 repository, download checkpoints/yue2_3b_bf16.safetensors (7.8 GB, or the 4 GB int8 version for lower-end cards) to models/checkpoints. Also download audio_encoders/sheetsage2_bf16.safetensors (1.4 GB) to models/audio_encoders.
3. Open Workflow, then Browse Templates, and select “YuE2 Text to Music”. The template defaults to the int8 version, so if you downloaded bf16, change the file in the loader node.
4. In the style field, describe the genre, vocals, instruments and tempo in English. In the lyrics field, paste the words with section labels such as [Verse] and [Chorus]. For Polish lyrics, type them normally with Polish diacritics.
5. Set the maximum length (I used 150 seconds). The model ends the song when it runs out of lyrics.

My settings were: “full” mode, temperature 1.0, top-p 0.95, top-k 100, repetition penalty 1.2, then 32 dpm_2 sampler steps with the sgm_uniform scheduler and CFG 1. These are the values from the official template.

I explain how to install ComfyUI from scratch in my ComfyUI from scratch guide. For more local models released in October, see the new local AI models roundup.

Which model should you choose?

• Polish songs for commercial use: ACE-Step 1.5. It was the most consistent in Polish, the fastest, and is released under the MIT license.
• Polish songs for personal use: YuE2 or ACE-Step. YuE2 had the best single attempt, but you may want to generate 2-3 versions and choose one.
• English: YuE2, which recognized 100% of the words in 32 seconds.
• MiniMax Music 3 is worth choosing if you specifically like its sound. It is the slowest and most memory-intensive. I cover it in more detail in my MiniMax Music 3 running locally article.
• DiffRhythm 2 is for English and Chinese lyrics only.

See how these models compare with Suno, Udio and Mureka in my AI music generator rankings.

The easiest option: Hangar

YuE2 is also listed in Hangar, my free manager for local image, video and music models on Windows. It needs no account or subscription and uses the ComfyUI engine. Hangar has the int8 version, which is smaller than the bf16 version I tested: the model is 3.96 GB and the sheetsage2 encoder is 1.39 GB, for 5.35 GB total. It is intended for lower-end cards, but I haven’t measured its speed or quality. The listing shows the CC BY-NC 4.0 license, which means you cannot sell songs made with YuE2.

1. Download the 14 MB installer from the Hangar page. It is not digitally signed yet, so Windows SmartScreen will show a warning: click “More info”, then “Run anyway”.
2. Hangar detects your graphics card (Nvidia with CUDA, from 8 GB VRAM; 12 GB or more is recommended) and shows whether the model will fit.
3. Find YuE2 3B in the catalog to see the file sizes, license and whether the model is allowed for use in the EU.
4. For now, generate songs in ComfyUI using the instructions above. Hangar can set up ComfyUI portable from scratch (it took 231 seconds on my system) or connect to an existing ComfyUI installation.

Hangar lists the model, but generation through Hangar is not yet available. As of October 17, 2026, you still need to use ComfyUI as described above. Read more about the tool on the Hangar page.

Frequently asked questions

Can YuE2 sing in Polish?

Yes. In my three attempts, Whisper recognized 84-99% of the words, averaging 91.2%. YuE2 may repeat a line or skip a section, so it’s worth generating 2-3 versions.

How much VRAM does YuE2 need?

The bf16 version used 13.3-16.3 GB of VRAM at peak on my system. The 4 GB int8 version should fit on a 12 GB card, but I haven’t tested that.

Can I sell songs made with YuE2?

No. The CC BY-NC 4.0 license prohibits commercial use. For music you want to sell, choose ACE-Step 1.5 (MIT) or DiffRhythm 2 (Apache 2.0); the latter is suitable only for English lyrics in this test.

How long are YuE2 songs?

Length depends on the lyrics and the limit you set. With 83 words and a 150-second limit, the songs came out at 2:07-2:18 min.

Is it better than Suno?

I didn’t compare the sound with Suno in this test. One advantage of a local model is that there are no download limits (Suno has had monthly limits since September 3, 2026) and you don’t pay for each song.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources

• m-a-p/YuE2-3B on Hugging Face
• Comfy-Org/YuE2 (ComfyUI weights, CC BY-NC 4.0 license)
• DiffRhythm 2 on GitHub
• My measurements on an RTX 4090, October 17, 2026: 13 recordings, Whisper large-v3.

Kinetyra