\u00b7 \u00b7 \u00b7<\/p>\n\n\n\n
Choreography\uff1aWhat a player actually does<\/strong><\/h6>\n\n\n\nThe choreography feature is four steps.<\/p>\n\n\n\n#<\/th> Step<\/th> What it is<\/th><\/tr><\/thead> 1<\/td> Pick a song<\/td> From the tracks in the game<\/td><\/tr> 2<\/td> Line up moves<\/td> Drag move cards onto a timeline<\/td><\/tr> 3<\/td> Set the stage<\/td> A stage theme, and the outfit your idol wears on it<\/td><\/tr> 4<\/td> Watch<\/td> A 15\u201330 second stage video of your idol doing exactly that<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\nWhat matters here is what the player is not<\/em> asked to do<\/strong>. <\/p>\n\n\n\n\n- There is no prompt box. Nobody types a description of a dance and hopes.<\/li>\n<\/ul>\n\n\n\n
That’s the whole bet of this project, in a single feature: the player is directing, not prompting.<\/strong><\/p>\n\n\n\n![\"\"](\"https:\/\/projects.etc.cmu.edu\/aidol\/wp-content\/uploads\/2026\/10\/Choreography-1024x381.png\")
<\/figure>\n\n\n\nMost of what we tried didn’t work<\/h6>\n\n\n\n
We tried four ways. Three failed, each differently and none subtly \u2014 wrong order, bodies coming apart, one run that came back as a blank screen. The fourth worked, and it was the plainest of the four. Camera control went the same way: the obvious approach failed, the unglamorous one worked every time.<\/p>\n\n\n\n
We’re holding the specifics for the December playbook. But there is a lesson from this week that costs us nothing to give away, and it’s the one we’d actually recommend to anyone generating video:<\/p>\n\n\n\n
\nNever accept anything that just worked once.<\/strong><\/p>\n<\/blockquote>\n\n\n\n\u00b7 \u00b7 \u00b7<\/p>\n\n\n\n
Lyrics editing<\/h6>\n\n\n\n
We’ve been generating song lyrics. They come back listenable but not natural<\/strong>. So, we started rewriting them by hand, and in the rewriting noticed something about our own product.<\/p>\n\n\n\n\nPicking a reference track from a list is multiple choice. Changing the words is authorship.<\/p>\n<\/blockquote>\n\n\n\n
That distinction is the difference between a player who used a tool and a player who made a thing, and we want our players to feel the latter one. Also, it’s more likely for them wanting to share the version I had edited myself.<\/strong><\/em><\/p>\n\n\n\nSo the lyric-edit step goes into the game.<\/p>\n\n\n\n
\u00b7 \u00b7 \u00b7<\/p>\n\n\n\n
Cost Breakdown<\/strong> (Oct 5 – Oct 9)<\/p>\n\n\n\nTool<\/th> Used for<\/th> Type<\/th> Cost<\/th><\/tr><\/thead> OpenAI Images<\/td> art assets creation<\/td> API<\/td> $4.65<\/td><\/tr> Pollo.ai: $399<\/td> video generation + pipeline test<\/td> Subscription<\/td> –<\/td><\/tr> OpenRouter<\/td> image\/video generation<\/td> API<\/td> $22.68<\/td><\/tr> Suno: $30\/mo *1, $10\/mo *1<\/td> Music prompt template tests<\/td> Subscription<\/td> –<\/td><\/tr> ElevenLabs: $6\/month *2<\/td> Backend probes: lip sync, gacha, joint test; SFX<\/td> Subscription<\/td> –<\/td><\/tr> Claude: $125\/month * 6<\/td> Docs, architecture, coding<\/td> Subscription<\/td> –<\/td><\/tr> Total<\/strong><\/td><\/td> <\/td> $27.33<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n\u00b7 \u00b7 \u00b7<\/p>\n\n\n\n
Weekly Progress Report<\/strong><\/p>\n\n\n\n
| #<\/th> | Step<\/th> | What it is<\/th><\/tr><\/thead> | ||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1<\/td> | Pick a song<\/td> | From the tracks in the game<\/td><\/tr> | ||||||||||||||||||||||||||||||||
| 2<\/td> | Line up moves<\/td> | Drag move cards onto a timeline<\/td><\/tr> | ||||||||||||||||||||||||||||||||
| 3<\/td> | Set the stage<\/td> | A stage theme, and the outfit your idol wears on it<\/td><\/tr> | ||||||||||||||||||||||||||||||||
| 4<\/td> | Watch<\/td> | A 15\u201330 second stage video of your idol doing exactly that<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n What matters here is what the player is not<\/em> asked to do<\/strong>. <\/p>\n\n\n\n That’s the whole bet of this project, in a single feature: the player is directing, not prompting.<\/strong><\/p>\n\n\n\n We tried four ways. Three failed, each differently and none subtly \u2014 wrong order, bodies coming apart, one run that came back as a blank screen. The fourth worked, and it was the plainest of the four. Camera control went the same way: the obvious approach failed, the unglamorous one worked every time.<\/p>\n\n\n\n We’re holding the specifics for the December playbook. But there is a lesson from this week that costs us nothing to give away, and it’s the one we’d actually recommend to anyone generating video:<\/p>\n\n\n\n Never accept anything that just worked once.<\/strong><\/p>\n<\/blockquote>\n\n\n\n \u00b7 \u00b7 \u00b7<\/p>\n\n\n\n We’ve been generating song lyrics. They come back listenable but not natural<\/strong>. So, we started rewriting them by hand, and in the rewriting noticed something about our own product.<\/p>\n\n\n\n Picking a reference track from a list is multiple choice. Changing the words is authorship.<\/p>\n<\/blockquote>\n\n\n\n That distinction is the difference between a player who used a tool and a player who made a thing, and we want our players to feel the latter one. Also, it’s more likely for them wanting to share the version I had edited myself.<\/strong><\/em><\/p>\n\n\n\n So the lyric-edit step goes into the game.<\/p>\n\n\n\n \u00b7 \u00b7 \u00b7<\/p>\n\n\n\n Cost Breakdown<\/strong> (Oct 5 – Oct 9)<\/p>\n\n\n\n \u00b7 \u00b7 \u00b7<\/p>\n\n\n\n Weekly Progress Report<\/strong><\/p>\n\n\n\n |