When AI Copies the Structure, Not the Look
A Japanese studio argues the next imitation wave is structural: a model reading a finished work for its staging, editing rules and aesthetic logic, then rebuilding it with different material. If that lands, tool fluency and mid-level craft stop being the thing you sell, and only two layers are left.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

Style transfer is the version of AI imitation everyone has already priced in. It is why there are LoRAs for every illustration style with a name, and why "I can generate in that look" stopped being a selling point somewhere around last year.
A Japanese studio, CreativeEdge, is arguing the next wave is a level up from that. Their term translates roughly to hyper-imitation, and their definition is specific: it is not copying a finished work, it is a model reading a finished work for its structure, technique, staging, design rules, editing method and aesthetic logic, then rebuilding all of it with different material under different constraints.
That is a much bigger claim than style transfer, because style transfer copies what a thing looks like. This copies how it was built and why it holds together.
Their forecast puts the point where it becomes routine at 2027. Treat the date as a hunch with a month attached rather than a measurement. The reasoning behind it is worth taking seriously even if the timing slips.
The skill layer that stops paying
Two of their observations follow from the definition, and both are uncomfortable for anyone whose income sits on a specific application.
The first is that fluency in professional software loses scarcity. If an agent can drive Photoshop, After Effects and Blender across a project without a human translating between them, then "knows After Effects" stops being a moat and becomes a preference. The apps do not die. Knowing them just stops being the bottleneck, and bottlenecks are where wages live.
The second is that mid-level craft commoditizes. When production capacity is massively supplied, price converges on the cheapest acceptable output. Not the best output, the acceptable one. Most professional work is bought at acceptable, which is exactly the band that gets squeezed first.
None of this is a reason to stop learning the tools. It is a reason to stop counting them as the asset.
Three kinds of maker, and the one people underestimate
CreativeEdge sorts makers into three types: the mass-and-imitation type shipping many one-off pieces, the continuity type building a large body of related work and accumulating a relationship with an audience, and the craft-and-experience type competing on quality that is hard to replace, live work, and physical presence.
The middle category is the one that gets underrated, and I think it is the one AI makes more valuable rather than less. A model can produce a single competent artifact. It cannot accumulate a relationship. Continuity is a time asset, not a skill asset, and time is the one input you cannot buy in a generation call.
That has an awkward implication for portfolios. Fifty polished pieces with no through-line are the mass category whether or not AI made them. The same fifty pieces organised around one world, one voice, one recurring cast, are a different business entirely, because the audience is following the through-line and the through-line is the thing that was never for sale.
The third category is real but narrow. Live performance, physical objects, work where the body is the medium. It is not a fallback for people who cannot compete on volume. It is a separate market with its own economics, and the people in it were never competing on render quality to begin with.
The idol case: the method is not the product
The same studio is running a second experiment that makes the argument concrete. They are building a fully generated idol group, with no real performers: faces, bodies, voices, singing, music videos, drama and promotional statements all produced with generative tools, and a human creative director controlling the persona, songs, footage, story and messaging.
Their framing is the part worth stealing. It is not "AI makes an idol." It is "a creator operates a fictional entertainer and uses generative AI as the production method." Those are different businesses with different risk.
Which is why their first rule is: do not sell the fact that it was made with AI. Generated music and footage are inflating fast, so technical novelty cannot differentiate anything, and the more attention the AI angle attracts, the more negative evaluation comes with it. Disclosure stays mandatory. It is just table stakes, not a hook.
Their proposed hook is the maker. The obsession behind the project, the reason someone would push this far, becomes a second story layered on top of the fictional one. That is a genuinely sharp move and it has a hard limit they also name: push the author too far forward and the character stops being a character and becomes the author's project, which kills the thing the audience was supposed to be attached to.
And the honest line at the end of it: nobody has demonstrated a fully generated idol sustaining a fandom the way a human performer does. The experiment is unproven. That is stated in the source, and it should be stated here.
Separate the brand from the assets
The most practical idea in the notes is a filing instruction. Split external brand value from internal production assets, and manage them differently.
External brand value is what the audience is attached to: the character, the world, the voice, the release rhythm, the promise about what shows up next. Internal production assets are the machinery: character bibles, reference banks, prompt and workflow archives, trained models, project files, the studio's own documented decisions.
Most small studios have exactly one of these and rent the other. Some have a beautiful production pipeline and no audience, which means every project starts from zero distribution. Others have an audience and no machinery, which means every project restarts the craft from scratch. The ones that compound have both, written down, in places they control.
This is also where a platform collapse becomes survivable. If the brand lives in your own bibles and the assets live in your own repositories, changing a model provider is a migration. If the character only exists inside someone else's product, you do not have a brand, you have a tenancy.
A short working checklist
- Name the layer you actually compete on. Volume, continuity, or craft. Pick one and stop pretending the other two are your strategy.
- Build a series, not a portfolio. A recurring cast or world is the cheapest moat available, and it is the one a generation call cannot copy.
- Keep the craft that needs a body. Live, physical, performed. It is a smaller market with less competition from anyone typing.
- Log the method, not just the output. The archive of how a thing was made is what lets you rebuild it when the tools move under you. We made the same case for prompts derived from your own archive.
- Do not lead with the method in public. Ship the work, disclose the tools, and let the tools be the least interesting thing about it.
Where this argument is soft
The source is a studio's membership radio series and an essay series, not research. There is no measurement behind the 2027 date, no sample, no test. The three-maker taxonomy is a framework someone found useful, not a finding.
The part I would defend independently is narrower: production skill is being oversupplied, and oversupply always moves value toward whatever cannot be mass-produced. Relationship and physical presence are the two candidates that survive that test, and they survived the last round of automation too.
The complement to this piece is what clients pay for after video commoditizes, which works through the same shift from the buying side rather than the maker side.