YuE 2 vs AceStep for Portuguese music: creativity vs pronunciation, LoRA struggles
Inevitable_Pen9043 · reddit · 2026-11-03
A music-generation community deep dive comparing open-source models YuE 2 and AceStep. YuE 2 produces interesting melodies but makes Portuguese pronunciation errors; AceStep nails Portuguese but the songs feel uncreative.
The author trained LoRAs for YuE 2 with aitoolkit with mixed results: past 1,500 steps the model goes out of tune; tweaking LoRA strength, text encoder and temperature gave inconsistent gains. They're asking the community whether people generate with AceStep and enhance with YuE 2 or vice versa, and how to improve either model's output.
More from Multimodal
- Creator: Seedance 2.5's control is so good that AI video is 'essentially solved' — Uncanny_Harry · 2026-11-03
- Watercolor Lord of the Rings image set via Midjourney --sref codes — sergeantsref · 2026-11-03
- 86.2 Agent Hours With Opus 5.5 Produced a Hyperpop Music Video — apples_jimmy · 2026-11-03
- Sume partners with HeyGen: 100 free video generations for first 10,000 users — HeyGen · 2026-11-03
- Open-source ComfyUI node turns instrumental tracks into sung songs with your lyrics — White_Dragon_0 · 2026-11-03
- 'SIX FINGERS': AI girl sings through every era of AI video, 2021 to now — ClaudiaSlop · 2026-11-03