Tegaki: Director Commentary

The short did not begin as an AI film

Tegaki is a repurposing of a feature screenplay I wrote around late 2023 to early 2024. The original protagonist was an overworked animation director who had moved from engineering into animation. Even after reaching the work she dreamed of, micromanagement and management and marketing demands stripped away her creative freedom. In that version, her drinking worsens, she cannot leave the dream even as her health fails, and she approaches death in a hospital bed. The successful independent filmmaking life was a hallucination.

The feature was also built around her friendship with an engineer who kept asking why she stayed in a dream that was destroying her, while continuing to support her. In the short, engineering and animation become a conflict about automation, AI, creative labor, and consent. Some character designs and backstory were also carried over from the original project.

Original CT timing animatic, 17 minutes 2 seconds.

The feature also carried the feelings from losing a project I had spent years developing. Understanding rights and organizational realities does not remove the pain of losing work into which I had poured my life. I could not secure the funding to make it, so the screenplay stayed parked for years. My own contradiction around AI later gave the title a new place to exist.

Using machines as someone who learned by hand

I am a traditional artist and a machine-learning engineer in my day job. I spent years learning anatomy, drawing, animation, screenwriting, storyboarding, and filmmaking. When Stable Diffusion and generative video systems appeared, I felt more than engineering interest. As an artist, I feared that the years of learning these foundations might become pointless. There were months when I barely wanted to make anything.

5RPS is a feature film that I write, direct, produce, and fund myself. It is about tech-bro founders scamming venture capitalists. Around a day job, I spent a year making a roughly two-minute trailer, paying artists, studios, voice actors, and composers from my savings. When backgrounds became a major bottleneck, I decided to use Gemini for some of them because the budget and time were running out. I was open about that use.

The trailer's reception changed around the single fact that AI had been used somewhere. A year of drawing, cleanup, coordination with other artists, and payment disappeared from view. I also heard about an animator I know who was blocked by people they knew after experimenting with AI and had to find new communities. I will not identify that person or their specific work here.

Tegaki is therefore not a film that settles whether AI is good or bad. I understand the anger and concern around training data, environmental cost, jobs, and what cheap generation does to creative labor. At the same time, I understand the technology, use it, and have been hurt by it. Living in both positions is at the center of this film.

A film that began in the car

When Json Cunanan told me Higgsfield was running an AI film festival, I imagined a one-day hackathon. Around mid-August, I planned to pause my work on 5RPS for two weeks and use the deadline to practice storyboarding. During the roughly hour-and-a-halfdrive, I used ChatGPT Voice as a transcriber and conversational notebook. I began with genre, then realized I already had Tegaki. I joined the old screenplay, the backlash I had heard through my friend's experience and my own personal experiences, and the experience of an animator I know into one short. I arrived with the beats, and Json Cunanan liked the idea.

The car has become one of my temporary writing studios, where voice conversation can become a working outline during a drive.

The one-day plan collapsed quickly. We initially misunderstood the rules as not allowing our own drawings, so we spent about two hours developing characters through generation alone. I also thought we could give the beats to Codex, turn each beat into a Seedance prompt, generate one video per beat, and assemble the film. But a beat is not a shot. Compressing its information, emotion, perspective, and rhythm into one clip made a rushed, incoherent movie.

At home, I tried asking Codex to write the screenplay. It was clean and logical, but emotionally empty. As I edited it, I realized I was rewriting almost everything, so I finished the screenplay myself. AI helped with formatting, translation, organization, and later decomposition, but the dramatic screenplay and direction still had to be mine.

I also built an Electron app that kept a rough board, manual direction, and its production text prompt together.

The screenplay had roughly eleven or twelve scenes, so I compressed it into six shorter scenes that could fit the schedule. We also tried rough 3D environments and rough storyboard frames, but generations followed their limitations too literally and became flat.

Json Cunanan kept testing and found clean background plates to be an efficient way to establish environments.

When I tried removing storyboards and moving by text alone, it was technically faster but the pacing became worse.

Make the animatic the single source of truth

That failure showed me that a storyboard is not merely an input image for a model. It is where direction happens. When information arrives, where a character stands, which way the eyeline and camera face, and how long a reaction lasts do not automatically become a film through text. I built an end-to-end animatic and matched every shot to its prompt and production information.

Subtitled pacing animatic from the later production process.

The animatic contained the preceding and following shots, duration, pauses, and performance intention. When footage was good enough, I incorporated it and adjusted the timing around it. Json Cunanan could work from a timed blueprint instead of abstract beats, then carry out generation, editing, and post-processing. As my day job became busier, this division let me focus on the judgments that required direction.

The method also returned to how I made films alone before AI. I drew boards, recorded temporary voices, added rough music and sound, and built a rough film from beginning to end. For Tegaki, I used ElevenLabs for temporary character voices so I could hear timing and pauses. The point was to build the whole film first, then replace rough parts progressively.

The screenplay established the drama, the storyboard established the visual storytelling, and the animatic established time. Prompts, plates, keyframes, and video models operated inside those decisions. Instead of asking for a movie, we could ask for this specific shot in this film. That was when generated footage became usable as film material.

Read the frame by hand, increase the attempts

Json Cunanan researched workflows, found the plate approach, refined prompts, and handled much of the generation and post-processing. He edited material, cleaned up pacing, assembled later cuts, and led the final music and edit. I owned the story, screenplay, rewrites, storyboard, end-to-end animatic, pacing, shot intention, reviews, and final selection.

For roughly the final three days, Json Cunanan was busy with his own work, so I reviewed generations, topped up Higgsfield credits with my own money, and generated replacements for missing or unusable shots. Sometimes I repaired timing with freeze frames rather than endlessly regenerating. After pushing the material as far as I could, I handed it back to him for remaining shots, editing, and pacing cleanup. We reviewed and revised the film repeatedly.

For music, I directed the emotion and placement. I described a muted piano quality, shared cinematic references, and said when a scene should have no music. Json Cunanan, with stronger formal musical knowledge, generated candidate tracks in Suno. I listened in the car and chose what belonged in each scene, including the empty spaces. He led the final music generation and integration into the edit.

The livestream, chat, notifications, translations, and harassment screens could not be left to the generation model. Exact text and timing show how quickly she loses control of the conversation. I built and animated the audience comments, chat, notifications, translations, and other UI by hand. This was traditional compositing and motion graphics inside a machine-assisted production.

AI did not remove the work of directing

At every stage, I tried to remove a step. Could beats become prompts? Could Codex write the screenplay? Could scenes become shots? Could I skip storyboards? Could 3D or plates reduce the drawing? Each time we removed too much human direction, the film became worse. The screenplay lost emotion, one clip per beat lost storytelling, 3D became flat, and pacing failed. A model can make astonishing material, but it does not automatically know what film I want.

The tools became more useful when I restored the foundations. I wrote the drama in the screenplay, cut shots in the storyboard, set time in the animatic, and judged camera, perspective, pose, acting, and continuity. I can recognize a wrong generation because I understand what the shot needs. I can recognize a beautiful clip that is three seconds too long because I understand the pause of a reaction.

My current position is not that traditional skill has become irrelevant. The better I understand screenwriting, storyboarding, drawing, animation, editing, acting, and music, the more precisely I can use the tools. Tools provide leverage, while foundational knowledge provides judgment. That is powerful for a tiny independent team, but it does not erase the questions around training data, the environment, employment, or the value of labor.

Whether AI was used does not explain the interesting part of a production. The useful questions are who made which decisions, what was delegated to a tool, and where a person took responsibility for the frame. In Tegaki, that exchange became both the production method and the subject of the film.

A one-day film, finished in roughly a month

Higgsfield extended the festival deadline by another two weeks, so what began as a one-day hackathon became roughly a month of production. We made it around two day jobs, and the extension helped enormously. Json Cunanan and I reviewed, decided, merged the final material, and delivered the finished film together.

Neither the subject nor the method returns to an easy conclusion. An animator who hates AI uses it to make a film about her experience. Inside that film she makes another film and submits it to an AI film festival. I made this film inside the same contradiction. At the end of the recursion is not a victory in an argument, but the act of finishing one film.

There are still places I could improve, and generated inconsistencies remain. Even so, I am happy that the film is complete, subtitled in English, and released for the festival. The film that was supposed to take one day made it here through roughly a month and the judgment of two people. I am genuinely happy that Tegaki is finished and released for the Higgsfield film festival.