TL; DR
- V1 cut down to seven steps, ending on publish or re-shoot? for playtest.
- The full month (four in-game weeks) is finalized ~40 minutes, day by day.
- Lip sync in generated video is fixed.
- The MV pipeline ran successfully end to end.
· · ·
Cutting V1 to one clean run
Quarters gave us a build worth showing and a clear read on what wasn’t going to be finished. So we scoped V1 down to a single uninterrupted run.
F1 Audition → F3 Home → F2 Interview → F7 Styling → F8 Recording → F10 Senior idol → your video → publish or re-shoot?What moved out:
| Cut | Where it went |
|---|---|
| F9 Expression class | V2 |
| F4 Dorm + F5 Quests | Folded into the Home scene |
| Everything after the video: publish growth, data report, ledger panel | Later |
· · ·
The 1 month in-game storyline:
One Month to Earn Your Debut.
You’re signed, but you haven’t debuted, and you have a month to earn it. Two things decide whether you do: your weekly evaluation grades, and how fast your following grows.
| How it ends | What it is | ||
|---|---|---|---|
| Week 1 | ~13 min | Being found, chosen, seen | V1 · the playtest build |
| Week 2 | ~7.5 min | From arranged-for-you to becoming someone | Next · music generation |
| Week 3 | ~8 min | Explore on your own; meet other players | |
| Week 4 | ~11.5 min | You stand on the evaluation stage and perform |
Week 2 is the one we build next, and it’s written down to the day:
| Day | What happens | |
|---|---|---|
| Mon | Writing your song | You brief the agency’s producer. It goes into production that takes a few days |
| Tue | Vocal training | Pitch-matching minigame. Your own voice, or a voice chosen to represent you |
| Wed | The song arrives | Two demos, in a generic reference vocal. You pick one. Only then does the backend render it in your voice |
| Thu | Dance training | A rhythm game charted from your own song |
| Fri | Expression training | Week 1’s memory game, with one difficulty step up |
| Sat | Posting | Pick one practice photo from the week, write something, post it |
| Sun | Weekly review | The producer’s report: this week’s grade and feedback |
· · ·
Lip sync had been drifting: 1.2 seconds of silence
Original assumption: the video model being imprecise.
It wasn’t.
Seedance does not track the audio you hand it. It listens to the audio, re-speaks what it heard, and animates its own rendition of that. So it was never syncing to the player’s recording. It was syncing to its own interpretation of it.
And what made it mishear: roughly 1.2 seconds of phone room tone before the player’s first word.
The fix is word-boundary trimming through ElevenLabs Scribe: cut to ±150ms around the actual speech.
· · ·
The MV pipeline
We ran the music-video pipeline end to end for the first time:
Pick a real K-pop MV and measure it. Write the concept. Generate the song in Suno. Build the storyboard on the song’s actual bars. Generate the character look and turnaround, then the set plates. Then generate the video, segment by segment.
Reference → Concept → Music → Storyboard → Look → Set → VideoThis is not only our backstage process. It’s the flow the player walks through in the game. Every step maps to a choice made in a UI.
· · ·
Cost Breakdown
| Tool | Used for | Type | Cost |
|---|---|---|---|
| OpenAI | Architecture, coding | API | $13.33 |
| OpenRouter | Concurrency probes ~$8; lip-sync probe $0.76; F7 collage probe ~$0.30 | API | $11.35 |
| Pollo.ai | Video generation | Subscription | $129 |
| Suno | Music generation | Subscription | $30 |
| ElevenLabs | Word-boundary trimming (ASR) | Subscription | $11.66 |
| Claude: $125*6 | Team efficiency | Subscription | – |
| Total | $195.34 |
Not approved cost: fal.ai ($2.93)
· · ·
Weekly Progress Report