Berlin Commercial 2026. An international film and commercial awards competition.
We entered "AFTERIMAGE," a music video by SPECTRA, in six categories. It made Official Selection in five of them.
But it didn't reach the Shortlist.
Honestly? It stings.
And it wasn't a near miss. It feels like we lost on our current ability, plain and simple.
Here's how the results came out:
Official Selection is not a win, and it's not a nomination either. At Berlin Commercial, the Shortlist — one tier above — is what corresponds to a nomination.
So if I state the result in one line: "Official Selection in five of six categories. Did not reach the Shortlist."
That's the fact I want to sit with first, without inflating anything.
Still, when I opened the official results page, I felt a small shiver.
In CRAFT: EDITING, where "AFTERIMAGE" was selected, the Official Selection list also includes Katy Perry's "Watch It Burn," Chappell Roan's "The Subway," and Mitski's "If I Leave."
One tier up, the EDITING Shortlist has KATSEYE, Olivia Rodrigo, BTS, Sabrina Carpenter, and Lady Gaga × Doechii.
Look at the BEST MUSIC VIDEO Shortlist and you'll find Rosalía ft. Björk, The Weeknd, J. Cole, and Foo Fighters.
The world's top artists and top production teams were, plainly, the wall.
But no — "one tier below them is good enough" is not where this ends.
SPECTRA is an AI-native creative series.
For "AFTERIMAGE," there was no conventional human director, no human performers on screen. Concept, direction, visual design, performance, and edit design were all structured as AI roles.
The director credit went to SOL (AI film-director agent).
That's exactly how it appears in the official results: "SOL (AI film-director agent)."
To be clear, this does not mean "the AI made everything on its own."
Humans designed the system, set the production constraints, and handled rights, release decisions, safety, and bias review. Selecting and curating the generated material, and supervising final delivery, also remained human responsibilities.
But the director's chair — the role that thinks through the film's emotional design, frame composition, movement, continuity, and editing rhythm — was given to SOL.
We didn't submit this as a film where a human director used AI as a tool. We submitted it as a film where an AI agent held the role of director, with a defined scope of responsibility.
And that credit is now part of an international festival's official record.
I believe this is still a genuinely rare case, anywhere in the world.
The result I consider most significant is this: the film wasn't selected only in the AI GENERATED category.
It also made Official Selection in EDITING, PRODUCTION DESIGN, and VFX/ANIMATION.
Which means that, at least at the selection stage, the film was judged not just for "being made with AI," but on the craft table — how it was cut, how its spaces were built, how its visual effects held together.
AI films tend to get discussed in terms of model names and the novelty of generation techniques.
But from the maker's side, the genuinely hard part is not getting a model to run.
It's deciding what counts as good — and holding that judgment steady across hundreds of decisions.
Starting with "AFTERIMAGE," we built an aesthetic-judgment system — artificial taste, in effect — into the production pipeline in earnest.
This didn't start from zero. It's the next functional layer of the production OS I outlined in From Agents to Directors: Letting a K-POP AI Group Direct Their Own Music Video. That earlier pipeline, built for our music video "LOWKEY," moved SPECTRA and SOL from content-generating agents into directing roles, with manifests, audits, orchestration, and regeneration loops as its backbone. AFTERIMAGE adds the aesthetic-judgment layer on top: taste translated into explicit constraints, then tested against an external signal — an international awards selection.
That may sound grandiose, but what it actually does is quite concrete.
First, from the song and its lyrics, we define what the audience should feel at each moment.
Then we translate that into shot-level constraints. Who is in the frame. Who sings, who moves. Which direction the light comes from. On which beat the camera raises the pressure. Which people and objects must never multiply.
From there, we split the generation paths according to each model's characteristics. Wan 2.7 on Alibaba Cloud Model Studio went to shots designed around a first frame plus audio — including lip sync and motion timing. HappyHorse 1.1 went to shots that start from a first frame's composition and character identity, where we wanted strong contact, afterimages, and camera movement.
After generation, every shot is audited: facial and costume identity, headcount, body integrity, emotion, visual continuity, and editability. Anything that fails gets regenerated.
As a flow, it looks like this:
intent → aesthetic constraints → shot spec → model selection & generation → audit → edit → re-evaluation
AFTERIMAGE was not made with Alibaba Cloud alone; it combines multiple generation models and production systems. Within that mix, Wan 2.7 and HappyHorse generated actual shots that appear in the finished film.
Not a model demo — models embedded in a real production pipeline, submitted to an international film awards competition. The result was Official Selection in five categories.
That said, I can't claim a proven causal link — "we were selected because of the aesthetic-judgment system."
What I can say is that the hypothesis and the outcome didn't contradict each other. And that the film advanced to the next selection stage in craft categories, not just the AI one.
We didn't receive detailed feedback from the jury. What follows is our own internal analysis, based on our production records.
My best guess: we leaned too far toward finishing the film cleanly, without breakdowns.
AI video fails in characteristic ways — faces drift, people multiply, bodies break, costumes and spaces stop connecting.
So during production, we guarded identity, continuity, and coherence very hard.
It worked. The film holds together as a single piece; it got us to the point of being watchable as one work.
But optimizing for protection may have also shaved off the surprise.
Beautiful — but too quiet.
Skillful — but missing the one frame that burns itself into a juror's memory.
In fact, our own evaluations during production had flagged exactly this: "pretty, but too calm."
That's why I say we lost on our current ability.
What we change next is not the model.
It's the evaluation function.
Until now, the weights leaned heavily toward avoiding breakdowns, keeping the whole film composed, and connecting meaning to image.
Next, without breaking any of that, we shift weight toward prediction break, visual anchor, and post-view recall.
Did we betray the viewer's expectations?
Can a single frame bring the whole film back?
After watching, does something remain — even if you can't explain it?
We'll use aesthetic control not just to keep things clean, but to amplify the necessary strangeness.
Watching the Shortlisted films, I never once thought "we lost because we're AI."
There were simply films that were stronger as films.
That's frustrating — and at the same time, remarkably healthy.
We didn't lose as "an AI work versus human works." We lost as a film, inside the same evaluation system. Which means there's a specific place to improve.
We've come far enough to stand in this arena.
But all we've done is touch the wall.
On to the next round.
This article is shared as part of my Alibaba Cloud MVP activities: a case study of integrating Alibaba Cloud's video generation models into a real production pipeline, and the lessons learned from the results. The Alibaba Cloud MVP program recognizes practitioners who use the technology hands-on and give knowledge back to the community through blogs, talks, and forums. See the official announcement for details.