Where we write up what we're building: AI beatmap generation research, product launches, experiments, benchmarks, and what we learn along the way.
We timed 2,501 ranked osu! mania songs automatically and compared the result with the mappers' own timing — ManiaGAI v2, ManiaGAI v1 and Mapperatorinator, at four tolerances. Where each method wins, and what the benchmark can't tell you.
DanceGAI is no longer only a generator: new maps now appear in Community, playable with their song and downloadable as .osz for their first 7 days. Here's how it works and why we built it.
How OsuGAI v1 turns a song into an osu! standard map: a gameplay representation of button presses and cursor positions, 118,000 training maps with osu!lazer Auto replays, four training rounds, the results — and the honest limits of the beta.
A full osu! mania mapset in about 20 seconds of GPU time instead of a minute or more: where the time goes in a ManiaGAI request, and the decisions — from the browser to batched difficulties — that took it out.
A look back at the ManiaGAI v2 release of October 1, 2026: what version 1 was, what v2 replaced, what changed for players in the weeks after, and what is still missing.
ManiaGAI v2 learned from 150,622 human-made osu! mania charts by about 14,500 mappers — far beyond the ranked section. What's in the corpus, by key count, difficulty and source, and why it was built that way.
Instead of mapping audio straight to notes, ManiaGAI first plans each section's gameplay with a diffusion model, then places notes with a MaskGIT-style transformer. The reasoning, the trade-offs, and what it makes possible.
