Level-5 streamed VISION 2026 II “Dream” on Thursday 10 September at 21:00 Japan time, and by the standards of a studio that spent most of the last decade quiet it was a generous show. Yo-kai Watch 2: Haunted Domain. Decapolice. Snack World: Reloaded. Holy Horror Mansion. A first look at Inazuma Eleven: Flames of Revolution, with a new lead called Nagi Furuhoshi. A date for Professor Layton and the New World of Steam — 10 December, on Switch, Switch 2, PS5 and PC. And held back for the end, a remake of Professor Layton and the Curious Village.
The side-by-side screenshots went up within hours. Backgrounds with that particular soft wrongness people have trained themselves to spot, character detail that dissolves when you look straight at it. Akihiro Hino posted a statement the following day, and the English pickups landed on 12 September.
He apologised. He also gave up nothing on the plan underneath, which is the half I want to sit with.
What Hino said the machine was for
The wording, in the translations that circulated: Level-5 wanted the presentation to be more spectacular, so it folded in some processing that doubled as an experiment with the latest AI. This may have offended people, and he was sorry for that. The studio would take it as a lesson.
Then the line he has been repeating for the best part of a year. Everything that makes a Level-5 game feel like one — scenarios, character designs, the basic shape of the world — is made by people, by hand. The machine is for the step after. Hino gave the example himself: taking artwork a human has already drawn and converting it into polygons, accurately.
Attached to that is a number. Level-5 wants a major title to go from around five years of development to around two.
A drawing is the wrong thing to convert
Take a Layton character. Flat colour, heavy silhouette, a nose you could hang a coat on, eyes sitting where a real skull would not put them. Easy to describe and miserable to build, because a face drawn in that style does not survive being made literal.
The nose reads as a nose from three-quarter front and as a spike from the side. So you do not model the nose in the drawing. You model a shape that resolves into that nose from the handful of angles the camera is ever allowed to reach, and you cheat the rest — pull the silhouette, break the symmetry, let the rig correct the head as it turns. The eyes get placed and scaled for the read rather than the anatomy. Rotate that head thirty degrees with nothing compensating and the whole face slides off.
It is the same argument as shading stylized hair as one mass instead of a hundred ribbons: you build a surface that lies in a controlled direction, because the honest version of it looks wrong on screen. The style lives in topology, in normals, in the rig, and in an unwritten tolerance list of camera angles.
An automatic conversion is faithful. That is the problem with it. It hands you the nose that is in the drawing, on a head that will be seen from angles the drawing never had to survive.
Five years is not spent making polygons
The other half deserves the same look. Where does half a decade go on a big title?
Not into asset creation. Modelling and texturing a cast is real work and it is weeks of it, but it is a known quantity, it schedules cleanly, and outsource houses have been absorbing it for twenty years. The five years go into the third version of the combat. Into the level rebuilt because the pacing was wrong. Into tone, into localisation across six languages, into certification, into the long tail of bug fixing where a studio finds out what it actually shipped.
Iteration is what costs the time, and iteration is expensive precisely because the decisions inside it are creative — the exact decisions Hino says stay with the humans. Drive asset conversion to zero and a five-year project becomes a four-year-and-ten-month project.
Which is worth having. It is not a two-year project.
Why it surfaced in a sizzle reel
There is a structural reason this landed in a showcase rather than in a game.
A marketing video is the one surface in a games company with a fixed public date, no engine to run inside, no QA pass, no certification, and nobody downstream who has to keep working with the file afterwards. A slightly melted window frame in a trailer breaks nothing. The same window frame in the game means an artist opens it, a lead signs it off, and it ships. The promo cut is where an experiment costs least, and it is also where every single person looking at the company can see it.
Level-5 had a hard date, seven or eight games to show, and a wish for the thing to look expensive.
What nobody outside the studio can check
Accounts differ on how far the AI reached. Some coverage describes artefacts in footage of the games; some points at the presentation wrapped around them, and reports of the broadcast note Hino’s presenter avatar restyling itself to match whichever game was on screen. His statement covers footage used within the event. He has been explicit that no AI-generated output data sits in the games that ship.
That claim is unverifiable from outside, and will stay that way until the Curious Village remake is in hands.
Worth remembering how wrong a crowd can be pointing the other way. Amplitude shot ninety seconds of live action with twenty-five actors and about a hundred and fifty crew, on built sets in Sofia, and was told a machine made it. An audience that has learned to spot artefacts has also learned to spot them where there are none. Level-5 is the case where the accusation landed and the studio said yes.
The Curious Village remake is a 2007 game whose whole appeal was that it looked drawn. It will be judged on whether it still does.