Source: raw/Forget_Suno._New_BEST_local_AI_music_generator_is_here.md (AI Search, YuE2 install tutorial, https://www.youtube.com/watch?v=9RtywbN—QE, uploaded 2026-09-15, about 25 min; sponsored by Luma) and raw/The_BEST_local_AI_music_generator_is_here.md (AI Search, MiniMax Music 3 tutorial, https://www.youtube.com/watch?v=aUcDyeZgz_k, uploaded 2026-08-15, about 22 min; sponsored by Higgsfield). Platform: YouTube auto-captions, fetched 2026-10-30. Caption fixes: “U A 2 / U2 / Ue2” → YuE2, “A Step” → ACE-Step, “Muse Scriber / Muse Scriptor” → MuScriptor, “JLM” → GLM.
Two open music models you can run for free in ComfyUI on a 16 GB laptop GPU or less. YuE2 writes an editable score (ABC notation) before rendering the song, which is why it can do cover songs, restyle a reference track and change key; the presenter calls it the best open music model available. MiniMax Music 3 is text-to-music only, up to 5 minutes, with a three-part description prompt. Neither is Suno quality by his ear, but both arrive as Suno tightened its download limits. The licences differ in a way that matters: YuE2’s model weights are non-commercial.
Key Takeaways
- YuE2 does covers; MiniMax Music 3 does not (yet). Feed YuE2 a reference track, choose melody-only or full-song reference, and give a new style prompt and lyrics: a pop ballad became jazz with brushed drums, Beethoven became theatrical hard rock, Jingle Bells went minor.
- YuE2 works by writing sheet music first. A “Generate ABC” node writes the vocal and instrumental notation, key and tempo, and the song is rendered from that score. That is what makes lyric, key and style edits easy.
- Agents can drive it. He linked YuE2 to an agent (GPT Astra or GLM) and edited a song by text: “keep the melody and lyrics, but make it more jazzy through reharmonization,” then “remove the guitar.”
- Hardware is modest. YuE2: a 7.8 GB BF16 checkpoint that fits about 8 GB of VRAM, or a 3.96 GB quantized one for under 4 GB; a song took just over 2 minutes on 16 GB. MiniMax Music 3: 9.8 GB full, 4.9 GB and 2.5 GB smaller builds, plus a 9.2 GB INT8 text encoder; a 1-minute song took 3–4 minutes.
- YuE2’s benchmark claim is self-reported: on a “song quality index” chart it beats MiniMax Music and ACE-Step 1.5, and even Suno V6 and 5.5, but sits slightly under Suno V5.
- Licence: YuE2’s code is Apache 2.0, but its model weights are Creative Commons non-commercial, per the presenter’s reading of the repo. He is unsure whether you can publish generations to Spotify for profit. This conflicts with an earlier wiki article that records YuE2 simply as Apache 2.0 (see the contradiction note on the MattVidPro article).
- Why it matters now: from September 3, Suno limited even paid Pro users to 20 downloads a month (Premier 60), commercial use requires a paid plan, and he says Suno adds watermarks or tracking to generations.
YuE2
Repo: github.com/multimodal-art-projection/YuE (the Python route); ComfyUI workflow JSON and models from huggingface.co/Comfy-Org/YuE2.
Install (ComfyUI):
- Update ComfyUI (
update/update_comfyui.bat). - Download the YuE2 workflow JSON and drag it onto the canvas.
- Download the audio encoder (1.4 GB) →
models/audio_encoders, and a checkpoint (BF16 7.8 GB, or quantized 3.96 GB for under 4 GB of VRAM) →models/checkpoints. Press R and select them.
Text to music (top half of the workflow):
- Style prompt: genre, style, pacing, instruments (“future bass, modern, energetic, inspiring”).
- Lyrics with meta tags: verse, chorus, pre-chorus, bridge, intro, outro.
- Max duration in seconds (default 360). Set it close to your lyric length; his first song ran long because he left it at the default.
- KSampler: seed (fixed reproduces the song; randomize for variations), 32 steps default, CFG up if lyrics or style are not followed.
- Decode: use the regular decoder with more than about 12 GB of VRAM (faster); use the tiled decoder below that.
- Instrumentals: say “instrumental only” in the style prompt and put the instrument description in square brackets in the lyrics field, for example
[steady build-up of staccato strings and ethnic drums].
Covers and restyles (bottom half): un-bypass the purple reference nodes (Ctrl+B), bypass the top Generate ABC and preview nodes, connect the reference branch’s output into Generate Music, select the audio encoder, upload the reference clip, choose melody only (for a cover) or full reference, and write the new style and lyrics. You can change the language of the lyrics too.
Demos he played: smoky jazz, 1930s boogie-woogie, Spanish flamenco, Russian folktronica, Korean emo rock.
MiniMax Music 3
Install (ComfyUI): Templates → search “music” → MiniMax text to music (or download the workflow from ComfyUI’s docs page). Download from the ComfyUI MiniMax Music 3 page:
- diffusion model: full 9.8 GB, a 4.9 GB build, or INT8 2.5 GB →
models/diffusion_models - text encoder (smallest INT8 9.2 GB) →
models/text_encoders - VAE (217 MB) →
models/vae
The description prompt has three sections (MiniMax’s recommended structure):
- Global metadata: genre, overall feel, pace, key (“lo-fi hip hop chill hop with jazzy extensions, laid-back and dreamy”).
- Vocal details: male or female, pitch, smooth or coarse, timbre.
- Arrangement: instruments, and what happens in the intro, verses, bridge and outro.
He suggests pasting the template into ChatGPT to write it for you.
Lyrics: meta tags (intro, verse, bridge, chorus, outro), secondary vocals in brackets, “mhm”s and “oh”s, and pauses all work. Max duration up to 300 s. Tiled decode cuts VRAM for long songs; turn it off on 24 GB for speed and quality. Batch size sets how many songs per run.
Demos: pop, punk rock, R&B neo-soul, progressive house, jazz, Chinese-language pop, metal, and country with a Southern accent. Verdict: “this ain’t Suno quality.”
Companion open tools he points to
- ACE-Step 1.5 XL: inpaint (edit sections of) existing songs or make music in the style of an existing song. The gap MiniMax Music 3 leaves.
- Foundation 1: generates individual loops from a reference loop, to stack into a full song in a DAW; tracks can be converted to MIDI.
- MuScriptor: upload a song and it separates instruments and transcribes each track, vocals included, to MIDI. Free Hugging Face Space; model sizes from a 5.5 GB large down to a 412 MB small (0.1B parameters). His test found a voice and an acoustic piano in seconds, with some note errors near the end.
Try It
- For background music on internal or draft content, install MiniMax Music 3 from the ComfyUI template and write the three-part description with an LLM. ^[inferred] Check its licence for commercial use first; the video does not state it.
- For a cover or a restyle of a track you have rights to, use YuE2’s reference branch with “melody only.” Do not publish YuE2 output commercially until you have read the weights licence yourself.
- Let an agent iterate. Give an agent access to the YuE2 workflow and edit by instruction (reharmonize, drop an instrument, change key) instead of re-rolling seeds by hand.
- Reverse-engineer a reference track with MuScriptor’s free Space to get MIDI you can rearrange in a DAW.
- If you depend on Suno, count your monthly downloads against the new 20-download Pro cap before a big campaign.
Implementation
Tool/Service: ComfyUI with YuE2 (workflow JSON) and MiniMax Music 3 (built-in template); ACE-Step 1.5 XL, Foundation 1, MuScriptor as companions.
Setup: models into models/audio_encoders and models/checkpoints (YuE2) or models/diffusion_models, models/text_encoders, models/vae (MiniMax Music 3).
Cost: free and offline. Disk about 5–10 GB (YuE2) or 12–19 GB (MiniMax Music 3) depending on precision.
Integration notes: YuE2 switches between text-to-music and reference-to-music by bypassing node groups; choose regular or tiled decode by VRAM. MiniMax Music 3 is text-to-music only.
Open Questions
- YuE2’s weights licence is the presenter’s reading (“Creative Commons non-commercial”); the exact CC variant and whether outputs are covered are not stated. The MattVidPro source says Apache 2.0.
- MiniMax Music 3’s licence is not discussed.
- The “song quality index” benchmark is not named or sourced beyond YuE2’s own chart.
- Who makes Foundation 1 and MuScriptor, and their licences, are not given.
- The opening line says YuE2 “even beats Suno V3” while the chart he later shows has it beating Suno V6; the V3 line may be a caption error.
Related
- YuE2 vs Suno V6 (MattVidPro) — the Pinokio install route and a head-to-head with Suno; carries the licence contradiction
- ElevenLabs Music V2 Review — a closed music model for comparison
- Local Video Models in ComfyUI — MiniMax H3 and LTX 2.5 — H3 can also generate audio, and pairs with these for music videos
- Local Image Models in ComfyUI — the same ComfyUI install pattern
- Claude Code Video Toolkit — where ACE-Step already appears in the wiki