Technical Polish Is Worth Five Points Out of a Hundred

Around 3,200 entries came in. TapNow selected 160 teams across Shanghai, Tokyo and Seoul, gave each one 32 hours, and asked for a short film. The Tokyo venue alone ran 45 teams, up from the 40 the organisers had planned, because the applicant pool was deeper than expected.

Most coverage of these events stops at the winner announcements. The part worth stealing is the scoreboard. TapNow's Creole Li gave the final-round weighting to AICU in writing, and it reads like a correction to everything the AI video scene has been optimising for.

The actual points

The final round was scored out of 100:

• Authorship, 25
• Originality, 25
• A story that depicts a relationship, 20
• Visual and sound expression, 15
• How deeply the theme is developed, and whether anything lingers, 10
• Technical execution, 5

Two categories about intent, one about human connection, one about craft, one about theme, and then a rounding error for whether the thing is technically well made.

Fifty people served as "Visual Ambassadors." Five of them watched every eligible entry start to finish and scored independently. The organisers told them not to weigh a team's reputation, follower count, or how expensive the tools were. So a team running the cheapest model in the room had no scoring disadvantage, and a team with a large audience had no advantage.

Li's written answer on what judges look for is blunt about the priority: "Above all, authorial expression and originality. What does the creator have to say about how connection happens, and have they found a fresh perspective or form instead of a familiar template."

What this does to a 32-hour plan

Take the rubric literally and the production schedule reorganises itself.

Half the score sits in two categories that no model supplies. A renderer cannot decide what you are saying about how two people fail to reach each other. It also cannot be original on your behalf, because originality is by definition the thing the model has already seen too many times. If you spend 26 of your 32 hours on upscaling, retiming and lip sync, you are polishing the 5-point line while leaving 50 points to whatever instinct you had at hour one.

The 20 points for "a story that depicts a relationship" is the most concrete constraint in the list. It is not asking for a plot. It is asking for two entities and a change in the distance between them. A montage of beautiful shots fails that test no matter how good the shots are. Two characters, one room, one thing that cannot be said out loud, and a final beat that lands: that passes.

The theme design did half of this work for the teams. The main theme was E.C.H.O., with four sub-themes: SAME FREQUENCY, OUT OF SYNC, ACROSS THE VOID, AFTERECHO. Each team drew two and picked one. That rule is a useful trick on its own. Drawing two forces a choice between two specific angles instead of one vague prompt, and the constraint arrives before anyone has fallen in love with an image.

Look at what won under those conditions. The overall bronze went to maru's Itterasshai ("Have a good day"), built on a mother tapping her son's back each morning before he leaves. The Best IP award went to ankimo's Re:ROAD. Japan regional awards went to Team ARS for TALK BACK, about a doll that talks back, and Team FUSION for KAERU, a six-minute science fiction piece about a robot and a cat on a planet that is not Earth. Small premises, one relationship each, no franchise scaffolding.

Where the rubric is weaker than it looks

A contest rubric is not a market. TapNow sells a production platform and compute, and a contest that rewards authorship is also a contest that separates its product from "anyone can generate a clip now." That does not make the weighting dishonest. It does mean you should read it as what a platform wants its ecosystem to look like, not as a universal law of client work.

Clients still buy polish. An ad agency will not accept a technically rough 30-second spot because the authorship was strong. The difference is that polish has become cheap and authorship has not, so polish stopped being a differentiator and turned into a baseline requirement. The rubric is what the field looks like after that flip.

There is also a measurement problem the organisers did not solve. "Originality" and "authorship" are scored by five people watching everything in a marathon session. Fatigue and ordering effects are real. A quiet, slow film placed after eleven loud ones reads differently than the same film placed first. Nothing in the published criteria addresses that, and no judging panel I know of has.

How to use this without entering a hackathon

Write the 25-point sentence before you write a prompt. One line, in plain language: what am I saying about how connection happens between these two? If the sentence is "a cyberpunk city looks cool at night," you have no authorship score and you are competing on the 5 points.

Then budget in the same proportions. Give the story and the shot plan most of your time, give the sound design real attention because it sits inside a 15-point category alongside the visuals, and treat upscaling as a final pass with a hard time cap. The teams that ran out of clock in Tokyo mostly ran out while fixing frames, not while deciding what the film was about.

Two more things worth copying from the format. Scouting was not a formality: teams were picked on portfolio strength, prior AI video experience, storytelling and visual expression, team completeness, and fit with the theme. That is a decent checklist for assembling a two-person crew on a paid job. And the draw-two-pick-one rule is easy to reproduce on a client brief. Hand yourself two candidate angles, kill one on purpose, and commit. The constraint is cheaper than another round of references.

If you are weighing model choice for a project like this, the cost logic in our Wan 3.0 breakdown still applies, and it is a budget question rather than a scoring one. What changed is the frame around it. The earlier piece on awards that judge process instead of output described contests asking how you made the thing. Arena Peak Asia asks a harder question: what did you have to say, and would anyone who has seen a thousand AI clips recognise it as yours.

That is the part of the 100 points no render farm can buy, and it is the part most AI video work still skips.