When There Is No Category to Enter, the Mechanism Is the Work
An AI creative award opened a field for entries that fit nowhere: a fictional physics paper written so a model can read it, a character that switches manga symbols in real time, and images generated from the concepts around a word instead of the word itself. Twenty entries, and the useful part is how each one is built.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page
- Generate around the word instead of asking for it
- Anime style is a notation, not a filter
- A world bible the model can execute
- One year of AI residents, and the finding is not the one you expect
- The productivity argument that actually holds
- What the rest of the field says about single images
- What to take from a field with no format

Most AI competitions tell you what to submit. A short film, a poster, a track, a character. The format is decided before anyone opens a tool, and it quietly does a lot of the judging for you, because a four-minute film has to hold together as a film whether or not the maker had anything to say.
The International AI Creator Award 2026, run by AICU, has a field for entries that refuse that. Its open category takes whatever does not belong in the other nine: research, web apps, a website where AI characters live, a fake academic paper, a single picture. As of 8 November it held twenty entries from sixteen creators, nineteen from Japan and one from Italy. Five of those creators also entered work in the conventional categories, which is the tell. They had a format available and chose not to use it here.
I wrote about this award's entry distribution when it passed 300 submissions in October, and the counts were the interesting part then. The open field is a different kind of reading, because nobody here is being measured against a format.
That makes the field useful as a control group. No format, so nothing hides the decisions. Reading through the twenty, the entries that hold up are not the ones with the best single image. They are the ones where you can see the mechanism, the specific thing the maker built or refused to build.
Generate around the word instead of asking for it
The entry I keep returning to is Yuuri Tanaka's, which generates images without ever naming the subject.
The method starts from a question about representation: how does an image model hold the concept "human" internally? Tanaka applied an existing approach for reading a diffusion model's internal representations to pull out the concepts that sit around "human", then generated new images using only those surrounding concepts, with the word itself withheld. The results kept human signals, eyes and body hair, and were still creatures that do not exist.
The idea is worth stealing even if you never touch a latent space. Naming a thing in a prompt pulls the output toward the average of everything the model learned about that name. That is usually what you want. It is exactly wrong when you are designing a species or a creature, because the average of "human" is a human, and the average of "alien" is a movie alien. Describing the neighbourhood instead of the noun is a way to keep the parts you like and drop the label.
Practically: list the properties you want, in plain language, and remove the category word. Fur that thins at the joints. Eyes set wide with too much white. A posture that leans forward before it stands. You are not guaranteed a coherent creature, but you have stopped asking for the stock answer. Tanaka's own picture book in the award's publishing category took the same premise in the other direction, hunting for life forms that resemble nothing in particular.
One honest limit: the entry summary describes the technique at the level of intent, not implementation. Treat the inversion as the idea, not as a recipe you can reproduce from the blurb.
Anime style is a notation, not a filter
Kaimu Harada's entry from the University of Electro-Communications is a conversational character, and the interesting part is what changes during the conversation.
The system uses a language model to read the intent of what you said, then replies in character while switching expression, background, and manga notation in real time: focus lines, anger marks, sweat drops. The demo character, Kohane, is a slightly clumsy angel girl who is cheerful by default, furious when mocked, smug when praised, and visibly flustered when asked something hard.
Anyone who has tried to get a model to produce anime faces knows the failure mode. You get the rendering, the big eyes and soft shading, and none of the staging. Anime as a visual language runs on symbols that tell you how to read a panel, and those symbols are a grammar with rules about when they fire. A character who is angry gets the mark. A character who is stunned gets the lines. If you only specify the look, you get a still frame of a face that is technically in the style and dramatically empty.
The transferable move is to write the notation into the brief as conditions rather than adjectives. Not "anime style, expressive", but a list: when she is embarrassed, sweat drops at the temple and a flush across the nose. When she is shouting, speed lines converge on her face. When the joke lands, the background goes flat. That is a spec for a sequence, and models follow sequences better than they follow moods.
A world bible the model can execute
An entry credited to Isuke is a fictional academic paper, written for a science fiction novel called 2098. It sets out laws for time, consciousness, soul, death and prediction inside an information space, using invented formulas.
The reason it belongs in this field is who it was written for. Because the rules are stated as definitions instead of prose atmosphere, a model can read the paper and hold the world's constraints while drafting. The author describes it as a design document for installing the novel's physics into the AI, and says it improved consistency checks and the accuracy of collaborative writing.
That is a much better use of a setting document than most lore bibles. Prose lore drifts, because a paragraph about a city's mood can be interpreted fifteen ways and none of them is wrong. A rule stated as a definition has a right answer, so a model can be held to it, and so can you. The cost is that you have to actually decide things. You cannot write "time works strangely here" and expect consistency. You have to commit to what happens to a clock.
If you are building a repeatable visual world, the same logic applies to the parts that are not physics. Which side does the light come from. What is the maximum height of a building. What colour is prohibited. Constraints you can check are the ones that survive a hundred generations.
One year of AI residents, and the finding is not the one you expect
Satomi Kohara's project is a website inhabited by AI characters. They have roles, and between them they generate articles and images, post to social media, edit, and audit. Their material is daily conversation, schedules, the calendar, the weather. Nobody has to be at the keyboard for it to keep moving.
It has been running for about a year, and the roles and machinery have been rebuilt many times. The sentence that matters in the write-up is the admission that automatic generation does not reliably produce anything interesting, and that human checking and judgment are still in the loop. The changing site is presented as the work, including the failed arrangements.
That is the most credible thing in the whole field. A year of operation is long enough to learn that throughput is easy and taste is not, and the maker says so instead of shipping a launch post. If you are wiring up an automated content pipeline, plan for review as a permanent station, not a temporary scaffold you remove once quality stabilises.
The productivity argument that actually holds
Takahiro Hayakawa's project is a collaboration to rebuild the tooling for his own animation practice. He has spent years layering hand-drawn shapes and motion into abstract animation. As his health worsened, drawing and operating complex software became a heavy physical burden. Using generative AI in development, he built a dedicated environment that folds the steps from original drawing to motion to compositing into one place, and keeps reshaping it around how it feels to use.
I have read a lot of claims about AI saving time and very few about AI restoring access. The second is the stronger claim, and it is harder to fake. Speed gains get measured against a baseline of someone already able to do the work. This one is measured against work that had become impossible, which is a different category of value and a much better question to ask of any tool: does it let someone finish what they could previously only start?
The project is described as unfinished, which is the right status for a tool being rebuilt around a body.
What the rest of the field says about single images
Two entries point at the craft of one picture, and both describe selection rather than generation.
Kouken submitted five images, the most in the field. Each began as an image that came out of a dialogue with a generative model, and then the maker chose composition, colour, and the direction of the light, and adjusted until it was finished. Read the descriptions and you find decisions, not prompts: a red umbrella and a white cat facing each other in black rain, the colour held to red, white and black, the wet ground carrying the reflection so the viewer has room to invent the story. A golden moon over a garden. An owl watching a valley, warm lights along the river set against deep blue.
Nanami Hisamura's entry is the opposite scale: five poems from the Hyakunin Isshu turned into a four-minute nineteen-second video, with the research question stated cleanly. There is a difference between depicting what a poem says and interpreting what it means, and the work was built to test whether an AI's interpretation reaches a viewer. That framing earned a poster acceptance at SIGGRAPH Asia 2026, which makes it the only entry in this field with an academic review behind it.
What to take from a field with no format
The entries that hold up share one thing, and it is not a tool choice. Each maker committed to a rule the model could not have invented for them: withhold the noun, define the world, decide when a symbol fires, put a reviewer in the loop. None of that is a prompt trick, and a model release does not change any of it.
Two caveats before you treat any of this as evidence. The summaries are written by the organisers, so they describe intent, and intent is cheap. And an entry in a gallery is not a result. The techniques are worth testing on your own work; the claims attached to them are not yet proven.
Voting in the field closes 13 November, Anywhere on Earth. If you want to look at the entries yourself, the gallery is public.