ACE-Step 1.5 turbo — full fine-tune

Every decoder parameter (1,575,458,880 of them) was fine-tuned on 11 Japanese songs, rather than training an adapter on top. Checkpoint: epoch 40, Adafactor, lr 2e-05, fp32 masters over two 16 GB GPUs.

Captions were trained with the tag jpdenpa in front, so keeping it in the caption is what pulls the fine-tuned style in.

Pre-rendered samples

Generated locally with these same weights (seed 42, 8 steps).

jpdenpa, japanese vocaloid pop, twin synthetic female vocals, playful hyperpop, chiptune stabs, call and response, satirical and upbeat

jpdenpa, japanese pop rock, powerful expressive female vocal, dense band arrangement, playful staccato phrasing, dramatic dynamics

jpdenpa, japanese pop punk, fast female vocal, playful vocaloid-style melody, electric guitar, punchy drums, energetic and comedic

jpdenpa, japanese punk rock, male vocal, aggressive guitars, rap section, gang shouts, sports anthem energy

jpdenpa, japanese anime opening, girls band, bright energetic pop rock, female vocal, electric guitar, driving drums, catchy uplifting chorus