YuE2 cover generation: per-song CFG/temp configs that beat the ComfyUI defaults
LilBrownBebeShoes · reddit · 2026-09-20
A detailed writeup on running the open-source music model YuE2 in ComfyUI for song covers — the default workflow settings don't work well, so the author tunes per song:
- Set both ABC Mode and Generate Mode to Full (not melody) when the lyric structure is unchanged — noticeably better results.
- Tune Generate CFG (1.2–2.0) and temp (0.80–0.95) per song, with topp 0.95, topk 100, reppenalty 1.20; DPM2 + sgmuniform sampler, 64 steps, Sampler CFG 1. Four cover examples across styles (2000s pop-punk, dream-pop ballad, post-grunge, piano ballad) come with full configs.
- A realaudio tokenizer NAR lora at weight 1.0 helps, and the author shares the LLM system prompt used for style expansion plus the full workflow file.
The model has lots of potential if loras can improve instrumental quality and add distinct voices — current vocals are somewhat generic.
More from Multimodal
- Experimental video mixes visual styles, generated with Seedance 2.5 on Adobe Firefly — umesh_ai · 2026-11-04
- Hailuo AI claims H3 scored perfect on language tests; Japanese creator praises its Japanese signage fidelity — Hailuo_AI · 2026-11-04
- YANAMI: 400K-dialog multimodal dataset shows models still fail at stateful multiturn interaction — Shengyu Wang · 2026-11-04
- Open-source Idol model turns a single image into 3D in 45s on 8GB VRAM — Promptmethus · 2026-11-04
- ComfyUI v0.44.0 released with new Patina texture nodes — Fcking_Chuck · 2026-11-04
- Krea 3 technical report released: a deep dive into data, architecture and training — bdsqlsz · 2026-11-04