Tegaki: Between Hand and Machine, Part 4

Turning fear into a film

Completed and released version. 1280×720, H.264/AAC, 24 fps, 10 minutes and 26 seconds. English subtitles are burned into the picture.

Tegaki began with fear as well as hope about technology. A friend in the anime industry had worked as a genga artist and sakkan on major anime series. After moving toward work with AI, they were ostracized by everyone they knew in that community. I placed that pain between Nozomi and Mia, then gave it images, performance, and time.

Within our partnership, I carried screenplay, storyboards, direction, and rough prompts. Json Cunanan built the agent harness and handled post-processing and automation. We created and produced Tegaki together. That is why we cannot treat hand and machine as opposing moral camps. Machine-learning automation increases the speed of repetition and the range of comparison. Animation knowledge decides what to keep, what to reject, and where a correction has to happen. The two worlds can help each other.

We made this film after-hours and on weekends, pooling our pocket money for production expenses and Higgsfield credits. With limited time and a small budget, we kept putting that money and time toward the next necessary attempt, correction, or export.

Working between those worlds, I turned that fear into a completed and released film with boards, shots, sound, and subtitles. We completed and released it together.

The completed and released submission project is available through the Film Festival Submission.

Knowledge turns automation into direction

Within the time available, the completed and released film still has improvement points. There are generated moments where we would make more hallucination fixes, and we cannot say that every inconsistency was resolved. Even so, we could not have reached a working film without foundational knowledge.

Screenplay writing let us decide what happens and when it has to happen. Rough storyboards fixed eyelines, camera side, poses, and prop continuity. The animatic placed those decisions in time. Even when the drawings were rough, shot length, pauses, and the order of reactions gave us something against which generated material could be judged.

A shot does not work when the camera and perspective are weak, even if the character is in the correct place. Smooth movement cannot carry an emotion when pose and acting are weak. Directing judgment comes first. Generation and manual correction test that judgment, return to it, and adjust it with finer control.

We corrected images manually to protect the meaning of the screen. We checked face direction, hand position, the state of the laptop, interface displays, logos, and audience continuity one by one. Foundations are not decoration from an older process. They are compressed judgment for deciding what to save when time is limited.

Restore readable screen states

The screen setup used flattened reusable components named chat_simple, chat_special, chat_box, donation, and donation_box. We started from two six-second source clips, one at 1280×720 and one cropped to 630×720. We placed an Instagram mark, a 64-profile audience grid, separately managed usernames, and planned bilingual chat copy. Usernames and private note contents are not reproduced here.

In the comparison between the generated and manually fixed passes, the generated screens are above and the fixed screens are below. We restored the blank laptop and interface, then corrected clip selection, chat and donation overlays, logo placement, and audience continuity. Within the time available, we prioritized readable screen states.

Generated pass above, manually fixed pass below. The source video has no comparison labels.

Build sound that follows the picture

We used Suno to search for short single-state materials. In Instrumental mode, we generated longer harmonic states, entered no artist or film names, and used the exclude-styles field to hold back unwanted directions. We auditioned candidates in the DAW, assembled them against picture, and cut what did not belong.

The early setup began with five cues conformed to the actual cut and a shared D axis. After auditioning, we found that a recurring four-note cell across four cues could unify the score, allowing each cue to have its own tonal centre. The movement from D to C to A minor to E follows Nozomi being pulled down, reaching the bottom, and returning above where she began. We listened to the picture and chose the scheme by ear.

We placed prepared upright piano, bowed glass and metal, sounds from the film's own object world, and tuned room tone small and deep in the mix. The flipbook riffle, pencil on paper, stylus on glass, keyboard, and export click follow what the hand is doing as timbre. Where music should not call attention to itself, we treated material as SUB_FLOOR beneath the room, so it would not compete with the screen's own sound.

On the title card, one prepared-piano D decays into the studio's sharp G hum. At the blackout, we used true digital silence. Around 7:52, at the first manual correction, E flat leaves and B natural enters. The moment a human hand corrects the machine appears at once in the image, the timbre, and the harmony.