TL; DR
- V1 is divided into F1-F13 (F stands for features)
- Unity project is setup in Perforce, and F1–F4 are built
- Backend skeleton is built: one doorway every AI call has to pass through
- Fallback plan for expensive things confirmed
- Weekly cost documented: $43.45 + CNY2798
· · ·
The map: F1 to F13
V1 is thirteen numbered modules. Here’s the whole journey in one table.
| # | Module | What happens | Status |
|---|---|---|---|
| F1 | Opening Storyline | 60-second intro: alarm → phone → street → recruitment pop-up | ✔ |
| F2 | Neural-Match Interview | 8-question dialogue producing a 16-cell archetype report and personalized invitations | ✔ |
| F3 | Welcome Kit | Two CEO offers, contract signing, and the kit that teaches the economy | ✔ |
| F4 | Idol Dorm | Home base — ID, wallet, quest desk | ✔ |
| F5 | Quest System | Debut quest, split into three ordered steps: look → music → film | |
| F7 | Styling Room | Generates the player’s exclusive 2D portrait | |
| F8 | Recording Studio | Records a ≤10s self-intro, in your real voice or a chosen TTS voice | |
| F9 | Expression Training | Expression coaching that raises camera confidence before filming | |
| F10 | Senior Encounter | A senior idol narrative beat — meet who you could become | |
| F11 | Publish Decision | Upload the finished video, or spend to re-shoot | |
| F12 | Data Report | Post-publish metrics, and the first dual-ledger entries | |
| F13 | Dual Ledger Panel | Compute ledger and Diamond ledger, side by side |
· · ·
What got build?
Frontend
- F1, F2, F3, F4 complete
- F5 menu in
Art
- F1 storyboard, first draft
- F3 and F4 UI assets
Backend
- Runnable skeleton up
- Video generation pipeline built
- Full chain running end to end:
- Illustration → three-view turnaround → player’s voice → video
· · ·
The video benchmark
We A/B tested video models on an identical brief — same prompt, same reference images, 12 seconds, 9:16, 720p.
| Model | Cost per 12s | Generation time | Verdict |
|---|---|---|---|
| MiniMax H3 Max | $1.35 | ~71 sec | Style, lip-sync and all four action beats on target |
| Seedance 2.0 | $3.64 | ~5 min | Same quality — 2.7× the cost, 4× slower |
| Seedance 2.0 mini / fast | $2.90 / $1.63 | Faster | Lighter quality, same engine |
| Seedance 2.5 | $5.64 | – | The most expensive, better quality |
| MiniMax H3 (6s clip) | $0.36 | 10–15 sec | The fallback plan |
The cheap won!
While our budget request is still pending approval, cost has become an important factor in our tool selection. For the current development phase, we’ve decided to use MiniMax H3 Max for stage performance videos and Seedance 2.0 for self-introduction videos. We’ll continue tracking their cost, quality, and performance throughout development to see how they hold up in production.
· · ·
Designing for Failure
Generation will sometimes be late, fail, cost too much, or be inappropriate for a particular player, so failure needs to be designed into the experience.
A fallback should be a version of the experience, not an apology for its absence.
| Scenario | Constraint | Designed Fallback |
|---|---|---|
| Video Isn’t Ready | Generation latency | Rough cut using existing audio, portrait, B-roll, and title package |
| Budget or Clock Is Tight | Cost / latency | 6-second Mini Max H3 generation path |
| Minigame Overruns | Development scope | Expression Training → Expression Quiz |
| Player Doesn’t Want to Use Their Voice | Player preference / accessibility | Four predefined TTS voices with no penalty |
| Server Dies Mid-Render | Technical failure | Persistent generation job; resume without paying twice |
· · ·
What it cost
Budget approval hasn’t come through yet, so this is still our own money.
| Service | Type | Purpose | Amount | Date |
|---|---|---|---|---|
| Fal.ai | API credit | Image and video generation | $10.70 | Aug 29 |
| ElevanLabs | Subscription | TTS — in-game content, voice test | $6.00 | Aug 31 |
| Fal.ai | API credit | Image and video generation | $26.75 | Sep 2 |
| Jimeng | Subscription | Video generation | CNY 2,798 | Sep 4 |
$43.45 USD + CNY 2,798
And from the fal.ai dashboard on Sep 4: $17.30 spent, $4.66 daily burn, 14 endpoints live.
Top drivers were Seedance 2.0 mini at $6.49, Seedance 2.0 reference at $3.64, MiniMax H3 Max at $2.23, and GPT-image-2 at just under $3.00 across two endpoints.
· · ·
Weekly Progress Report
Meeting Notes:
Needs changes:
- F1 needs to lose the real world: drop the player straight into Aidol World, and lead with the thing the game does best: voice.
- Urgent demo: video recording us and turn into AI video.
- Maybe, players can buy a better AI video generation model in game.
- A better in-game phone means better video quality; a monetization line
- Players can choose a minigame or a story video during the wait for AI generation.
Still arguing:
- The archetype question. Why does everyone look the same? Are we trying to set a beauty standard?
- “Is it too feminine?” Can male players play the feminine character doing feminine poses?
- What breaks at 100,000 players? How do we keep the game fresh?
Also on the list: Amazon Music and AWS visit CMU Sep 16–18.