正在加载视频...

视频加载失败

yeah this is peak

48,829 次观看 • 2 天前 •via X (Twitter)

16 条评论

Jaroslav Beck 的头像
Jaroslav Beck1 天前

Back then, we deliberately chose to build Beat Saber beatmaps by hand, as the quality of generated content simply couldn’t match it. Human touch will always have its own magic, and I wouldn’t give up on that. But I can see the tech getting to a point where I’d actually want to test some generated maps based on my preferred style.

nikhil 的头像
nikhil2 天前

oh this is sick! i have an eval suggestion: can you put in something like kanaria - king and see if it does something reasonable for the "left side, right side" part? they messed this up in pump but it's quite good in maimai

Denyah_ 的头像
Denyah_1 天前

what llm are you using? for a start this looks scarily promising

lyra bubbles 的头像
lyra bubbles1 天前

which am i using to help build/train this or which am i finetuning for this?

Denyah_ 的头像
Denyah_1 天前

Both!

lyra bubbles 的头像
lyra bubbles1 天前

the model itself is a qwen3 0.6b with a *lot* of conditioning and audio input layers stacked on (MERT, custom audio projection layer) and heavily tuned. chose this over a pretrain from scratch since i can inherit the attention patterns qwen learned in pretraining, plenty of useful geometry can carry over to mapgen. music adherence (where all the custom model design comes in) is still its weakest point, idt the current iteration will be what ends up shipping my collaborators are multiple instances of both fable 5.1 and gpt6 astra. one of the sessions iterated for 6 days mostly unsupervised, most others were heavily collaborative - i prefer being in the loop and guiding design decisions

lyra bubbles 的头像
lyra bubbles1 天前

there might be a case for scaling to up to 3b or something, but the llm capacity doesn't seem to be the limiting factor and id rather improve other parts of the stack before scaling the actual generator

Denyah_ 的头像
Denyah_1 天前

Yep makes total sense. strong foundation will really push this forward and tbh given how its trained for ~6 days it looks like great progress. good shit!

lyra bubbles 的头像
lyra bubbles1 天前

most of the actual training time has gone into testing myriad different RL approaches to see if i can comfortably hill climb a whole suite of metrics first, the actual training for the checkpoint in the vid took ~2.5 days across 3 runs. turns out the hard part is r&d still

TooCooFooU 的头像
TooCooFooU1 天前

have u seen this

rachel 的头像
rachel1 天前

WOAH this map feels crazy readable. insane results

lyra bubbles 的头像
lyra bubbles1 天前

just wait it'll get better :D

小俞- 的头像
小俞-1 天前

What is this fruit fly’s name?

dz 的头像
dz1 天前

RELEASE WHEN?????

normal nice quinn 的头像
normal nice quinn2 天前

omfg!!!! excited about this

Harper 的头像
Harper1 天前

Check dm's!

相关视频

This is so peak, why is he so peak
0:31

Sensitive content

This is so peak, why is he so peak

aeeraa.

83,384 次观看 • 1 年前