From intent to an
executable character
The character engine behind our virtual broadcasting system
We are building a total system that connects broadcast preparation with live execution and streaming and audience interaction
Its character foundation turns an illustration or a written goal into decisions that can be executed and measured
Understand the goal
A visual reader extracts character attributes from illustrations
The decision policy combines that input with a written goal and the context of earlier choices
A frozen Qwen3-VL-4B backbone provides shared visual and language representations
LoRA adapters and a choice head learn to score the available candidates
Compose an executable program
The policy chooses a body and face composition together with an outfit palette and body proportions
It then selects the main hair and additional hair groups plus expression and color
Each choice belongs to a typed candidate set
The executor assembles the selected components into a character and measures the result
Separating decisions from execution lets the system reuse reliable operations while learning which combinations best express the creator’s intent
Learn from execution
Oracle search
A search process scores candidate programs and executes the strongest alternatives
Measurements select the training target for each goal
Supervised decisions
The policy learns a distribution over useful choices
Soft targets preserve near-equivalent alternatives instead of treating every other candidate as wrong
Reward-driven refinement
The reinforcement-learning implementation samples programs and scores their results
A reference policy constrains the update and periodic full execution checks the fast reward
Measure the assembled character
The evaluation checks attachment and seams as well as penetration and exposed scalp
Appearance matching and composition diversity contribute only after the geometric gates pass
Training and validation split character lineages so alternate versions of the same source cannot silently cross the split
Engineering snapshot
The engine specification dated 11 October 2026 records the following development state
| Component | State |
|---|---|
| Visual reader | Initial stage trained |
| Reader extension | Expression and proportions training |
| Oracle | Label generation in progress |
| Decision SFT | Training preparation |
| RL | Implementation complete |
| End-to-end evaluation | Planned |
The initial reader’s reported lineage-separated validation accuracy is 0.714 with expected calibration error of 0.056
The next evaluation compares search and learned policies on character validity and intent matching as well as human preference
From the engine to the broadcast
This character engine is the technical foundation for a broader service
Our product direction connects preparation and performance with streaming and audience communication so creators can work through one coherent system