YuE2 Studio 3.8.4

2026-11-05 — 3.8.4

Fixed

• The second section and the chorus no longer come out as sound-alike nonsense on newer NVIDIA cards (#194): the 3.8.0 move to a newer ggml brought a CUDA flash-attention kernel that is numerically wrong on Blackwell-generation cards (compute capability 10.0 and up, such as the RTX 5060 Ti), so a song sang the words right at the start and drifted into "sounds like singing but means nothing" later. The studio now turns flash attention off automatically on these cards; other cards are untouched and their speed does not change.
• Songs are no longer lost to an out-of-memory crash on cards of 8 GB or less: the smaller key-value cache that kept AMD cards of 8 GB or less stable was never applied to small NVIDIA cards, which crashed with an illegal-memory error mid-song (seen across many crash reports). The cap (9216) now covers any card of 8 GB or less; a cache size set in Settings - Engine still wins.
• A finished song is no longer lost when the system temp folder cannot be written: the FLAC was written through a temporary file, and a temp folder gutted or blocked by Windows lost the song at the last step. It is now encoded in memory.
• Karaoke times original songs whose voice is buried under the mix (#192): when nothing is heard in the mix and the vocal separator is installed, the voice is lifted out and timed on its own, so a freely generated original is timed where the mix alone reported nothing - now with cloud recognition too.
• Exported karaoke subtitles end when the line is sung, not at the next line (#189): SRT and LRC leave instrumental stretches empty instead of holding the last line, and the timings line up with the song.
• Your chosen maximum duration is kept when a new Create loads a request (#184): reuse, an example or the writing assistant no longer resets it to the default.
• The autogenerated lyrics fall back to your language instead of Russian (#198): a text-language selector was added, and under "Automatic" the lyrics follow the studio's own language when the request does not settle one.
• Trained LoRA checkpoints show up again (#177): the studio now also finds checkpoints saved under the trainer's newer names (*_joint_{ar,nar}_step<N>), not only the older native-{ar,nar} ones.
• Each changed component on the update screen has its own line (#180) instead of repeating the HT-Demucs description.
• A score whose M:/L:/K: header is dropped is sung instead of thrown away: the engine fills a default time, note length and key as a last resort.
• The Russian typo in the score-edit assistant's placeholder is fixed (#195).

Added

• A 1980s Eurodisco LoRA is in the adapter catalogue.