The AI Idol's Real Problem Is That It Has No Past

A creator studio in Japan has been running a public experiment: build an AI idol group from scratch and document it. The second session, streamed on October 22, reported something more concrete than concept art. They had a character who could sit for a 56-second interview and look like she had been alive before the camera turned on.

The method was not a better prompt. It was a life.

One month of biography, written with two coding agents

The character is called Saki. Using Codex and Claude Code, the team wrote out her timeline from birth to the present: where she grew up, what her family was like, what she failed at, what she kept doing anyway. They call the resulting document an "emotion engine," a term they coined for the project, and they use it to condition how she behaves on camera.

The detail that matters is the time cost. About a month of work, and explicitly not something you can hand to an agent and walk away from. The agents were useful for producing volume, structure and consistency across a long document. Deciding which events shaped her, and which ones to throw out, stayed with the people.

That ratio is the honest part of the report. Most conversations about AI characters start with the model and treat characterisation as a prompt prefix. This team inverted it: the model was the last decision, the biography was the first.

Then the technical filter: skin

Once you have a character, you find out that current video models disagree about what a face is.

The team's requirement was specific: facial muscles and skin detail. Not overall image quality, not camera language, not motion. Skin. They tested the field and reported that only Seedance 2.x and Wan 3.0 cleared the bar. Kling 4.0 was dropped for skin detail, and MiniMax H3 showed the same problem. All the Saki footage, including the interview and a 1:25 live-action-style version of her podcast, was generated on Wan 3.0. They skipped Seedance 2.5 entirely and put the cost at roughly a third.

Two things follow from that, and neither is "use Wan 3.0."

First, model choice is per shot, not per project. Skin holds the close-up. A wide shot of a street does not need the same model as a face three seconds from the lens, and paying the premium model rate for the wide is how budgets disappear. The cost-lane argument is the same one, applied to a specific body part.

Second, the thing that breaks an AI face in 2026 is no longer composition or lighting. It is the micro-texture of skin and the way facial muscles move when someone is holding back a reaction. Those are the shots that decide whether an audience reads a person or a render, and they are the shots to test first when you are evaluating a model for character work. Test the hardest frame, not the hero frame.

Limited animation, redefined

The other half of the project is the music-video track, and the team's notes there are a useful vocabulary for anyone generating "moving illustrations."

They are not trying to make anime. They are trying to redefine limited animation for generated footage, with a list of rules that reads like a set of constraints for a prompt:

• A still image is the base, not a compromise
• It should not move smoothly, and should not try to
• It should not drift into a 3D CG look
• If you pause it, it still has to work as an illustration
• The originality has to come from exaggerated visual expression, not from motion fidelity

Those rules exist because the failure mode is obvious to anyone who has tried it: you ask for motion, the model animates everything, the illustration dissolves into a smooth, weightless, faintly plastic clip. Holding the frame still and animating two things is harder to brief and much better to watch.

The business claim, which is the weakest part

The studio's strategic argument is that an AI idol's biggest weakness is having no life, and that selling the technology is the least likely route to success. They also argue that jumping straight to an "AI idol debut" is a recognisable failure pattern, and that the starting point has to be a format that already has an audience.

Their example is a podcast, Saki's Showa-era street walk, which then got a live-action-style video version. That is a sensible order: audio first, because a voice with opinions is cheap to produce and easy to test, then visuals once people already want to hear her. The music-video route was their first instinct, and they reported that a music-plus-illustration video alone was too weak to serve as an origin.

I would treat this part as a hypothesis, not a finding. The report is a progress note from a paid workshop series, not a revenue result, and the numbers that would settle it (retention, merch conversion, subscription churn) are not in the public post. What is public is the production method and the model comparison.

What you can copy this week

Write the timeline before the character sheet. Not a backstory paragraph. A dated list of events with consequences: age, what happened, what she decided about people afterwards, what she will not do. Then derive the visible things from it. Speech habits, posture, the clothes she would actually own, the topics she avoids. The test is simple: could a stranger read two paragraphs and predict how she reacts to being interrupted?

Keep the canon where an editor can reach it. The reason a long timeline works as an "emotion engine" is that it turns every shot decision into a lookup instead of an invention. Why is she looking away here? Because of an event at 19. That is much faster than re-deciding her personality at each prompt, and it is the mechanism behind consistency that the text-first character method describes from the stills side. The sequence-level consistency guide covers the reference and continuity plumbing around it.

Do not let the agents decide what matters. Codex and Claude Code are good at expanding a sparse timeline into a coherent, contradiction-free document, and they are good at flagging where two facts conflict. They are not good at knowing which single event deserves to be the reason a character flinches. That judgment is the month of work.

And check the skin before you fall in love with the motion. A clip that reads as a person for three seconds and as a render for the next ten is not a character problem. It is a model choice you can measure in an afternoon.