I Tested Maestro AI Song and Music Generator: Free AI Music Generator on Your Own PC

Plus:
– uploading a source song
– recreating a trending track
– a male and female duet
– gospel choir
– instrumentals
– music LoRAs

📥 All songs from this video: link to Skool classroom, coming soon

Chapters
0:00 Every music model, one brief
0:19 The brief and the rule
0:56 ElevenLabs Music (paid, for comparison)
1:21 The music panel
3:09 ACE-Step models, the LM error, Advanced settings
5:18 ACE-Step SFT and LM versions
6:58 MiniMax-Music3
8:08 YuE2
8:48 HeartMuLa: download, error, seed
11:28 Generation times
11:42 Upload source song
13:13 Melody only: cut off at the limit
14:24 Fixing the ending
15:22 Copying a famous artist
18:03 A famous song from one upload
18:43 Two singers
19:54 Gospel choir
20:53 Instrumental
21:47 Music LoRAs
24:40 Train your own LoRA

Join our free AI filmmaking community on Skool:
👉 https://aishotstudio.com/skool

🎵 Maestro: https://github.com/Blizaine/Maestro
⚙️ Pinokio: https://pinokio.co
🎛️ Free YuE2 style LoRAs: https://huggingface.co/pshields/yue2-style-loras

Generation times on an RTX 5060 Ti 16 GB:
ACE-Step XL Turbo 4B: 7 s
ACE-Step XL SFT 4B: 27 s
ACE-Step XL Turbo LM_4B: 28 s
ACE-Step Turbo LM_4B 2B: 33 s
ACE-Step XL SFT LM_4B: 49 s
YuE2 3B: 52 s
HeartMuLa RL OSS: 1 min 52 s
MiniMax-Music3: 4 min 6 s

Errors in this video:

– ACE-Step XL Turbo 4B: “ACE-Step 1.5 model mode 1/2/3/4 requires an LM definition (text_encoder_URLs).” Fix: open YuE2, set Composition planning to Melody and chords, switch back to ACE-Step. A page refresh clears it too.

– HeartMuLa: “Generation failed”, and in the terminal: “HeartMuLa does not support reference audio yet.” Cause: a song uploaded earlier for YuE2 was still attached. Fix: refresh the page.

– Importing a slider LoRA: “[WinError 32] The process cannot access the file because it is being used by another process”. Real cause: the format is not supported; Maestro needs a full combined LoRA.

– Training a LoRA: “Sound adaptation currently requires at least 20 GB VRAM”.

Bugs in this video (no error message):

– KeyScale asks for a key like “C major, F# minor”, but the field only takes numbers. A viewer’s tip, not tested yet: it may want the pitch-class number (C 0, C# 1, D 2, D# 3, E 4, F 5, F# 6, G 7, G# 8, A 9, A# 10, B 11). How it marks major or minor is unknown.

– ACE-Step models show an “Audio In” badge, but the Music panel has no upload for them: only YuE2 takes a source song.

– Write Song fills HeartMuLa’s style field with a long paragraph, but the field asks for short tags.

– The pencil loads a song’s settings back, but Composition planning comes back as Direct generation.

– YuE2 Melody only with a source song: every song ran to the limit and was cut.

#maestro #localai #freeaitools

🔧 The Maestro songs in this video were generated locally with Maestro, built on WanGP: https://github.com/deepbeepmeep/Wan2GP