ACE-Step

/

Ace-Step1.5

Like 895

Follow

ACE-Step 1.32k

I've been running ACE-Step 1.5 locally for a few weeks (acestep.cpp, Vulkan, Radeon RX 6600 8 GB, Windows 11) and wrote up what I learned, in German and English:
https://huggingface.co/spaces/charlysporty/lokale-musikerzeugung

A few findings that might be useful to others:

• Every melody needs a lead. Instrumentals without a named lead scored 100 % on my scale-adherence metric but sounded "not really harmonious". With a lead instrument they were fine – but only if the lead instrument comes first in the caption. Otherwise the first instrument named in the template takes over (e.g. the harmonica in a Chicago blues caption beat the guitar I had chosen).
• Melody guidance via the cover task: audio_cover_strength 0.3 with a deliberately plain, dry synth source gave the best tone and phrasing; from 0.5 the synth timbre bleeds through.
• Blind test vs Stable Audio 3 (both instrumental, same lead, seed, tempo and loudness, 5 styles): the listener preferred ACE-Step 4 times, once a tie, Stable Audio never – although all ten versions were rated "very good".
• Performance: XL-Turbo renders 120 s in about 210 s (up to 6.5 GB VRAM), an 8-minute song in about 17 minutes.
• What did not work: register words ("low register"), mix words ("subtle bass"), negations, and the complete task with the base model (phantom vocals).

Feedback, counter-examples and your own measurements are very welcome – especially whether the lead rule holds for other listeners and hardware.
(Programming, measurements and write-up were done with Claude as an AI assistant; listening judgements are mine.)