NEXUS AI Studio
A character IP engine that turns one seed line into world, scenes, video and chat
A studio engine that builds character IP end to end with an in-house anime engine, a semi-realistic engine and a self-hosted chat model. Every match in ondo, the latest work, was generated here. Eight engines mesh through pipelines, judges, serving and review contracts.
Visit NEXUS AI Studionexus-aistudio.vercel.app
- Role
- Product · architecture · engineering
- Contribution
- 100%
- Year
- 2026
- Platform
- Generation engine · web · 5 languages
108 characters had to keep the same face across 10,000 scenes
Any model can make one image or one chat turn. Keeping 108 characters on the same face and voice across 10,000 scenes, with predictable cost, is hard. So I built the judging and cost control into an engine: a process you can operate.
- Point 01
No improvised picks by the LLM
Code assigns composition and lighting deterministically from tables. Scene titles named after objects dropped from 60% to under 20%.
- Point 02
A judge that makes it redraw
The semi-realistic engine measures every result with face embeddings and redraws anything below the bar. Face identity 0.458 → 0.738, processing time 71s → 31s.
- Point 03
Predictable cost
Marginal cost per character $0.54–0.58, fixed infra $57 a month. Forecast and actual spend stay within 6%.
- 01
World bible
One seed line sets stage, tone, relationships and desires
- 02
Season arc
Events per beat, with setting and rating hints per beat
- 03
Scene design
Captures a moment by place, emotion and action
- 04
Direction axes
Composition and lighting assigned from tables
- 05
Image · Video
Designed scenes become images, then 5s clips
- 06
Chat · Distribution
Nexus Talk chat, auto-listed on 4 external platforms
Results
- Characters
- Designed scenes
- Video clips
- Cost forecast error
- Engine sitenexus-aistudio.vercel.app
Nexus Talk →
A 27B Korean model that talks with human turn length and rhythm