# Technical information

This record concerns the 45.29-second, 14-shot main film shown after the opening free talk and title, not the entire making-of video. Model names are production-time records, not current recommendations.

## Finished specification

- Duration: 45.29 seconds (45.28 seconds in the build record)
- Shots: 14
- Video: 1920×1080, 24 fps
- Audio: yes
- Live-action source: 30 seconds of night trail-camera footage

## Technology by stage

| Stage | API, model, or tool | Purpose |
|---|---|---|
| Story and staging | OpenAI Codex | Organized the cause-and-effect chain from live action through “Shh,” approach, tap, antler drop, surprise, apology, and retaliation |
| Excel storyboard | Codex, Python, openpyxl, Pillow | Managed 14 shots, reference images, antler side, distance, prop, and posture |
| TAKO reference | Shared character ledger | Fixed the round goggle head, slim metal body, and straw hat |
| Keyframes | OpenAI Image Generation | Designed approach, one-antler continuity, upright TAKO, deer backing, and face reaction. The model ID was not recorded, so none is guessed here |
| Generated video | Google Vertex AI Veo 3.1 Standard `veo-3.1-generate-001` | Used first/last frames and short causal segments to create trail-camera-style motion |
| Japanese TTS | Google `gemini-3.1-flash-tts-preview` | TAKO’s “Shh” and “Sorry.” The deer’s dialogue was removed from the final version |
| Three trajectories | Deterministic coordinate compositing in FFmpeg 7.1 | Moved exactly three pellets, one at a time, from the deer to TAKO’s face and fixed all three contact points |
| Sound, subtitles, final edit | FFmpeg 7.1, ASS, real ambience, local sound effects | Discarded API-generated audio and synchronized timestamp, dialogue, footsteps, launches, and impacts |
| QA | Codex, FFmpeg, openpyxl, human review | Checked one-antler continuity, contact, posture, three start/end positions and times, and the final cut |

## Connecting live action and generated video

The source was not treated as a loose visual reference. Its fixed framing, monochrome infrared look, timestamp, forest ambience, and noise became an environment ledger. The generated sections matched those properties and were connected through short segments and first/last frames.

## Why one long generation was avoided

Asking one generation to remove an antler, turn the deer, back it up, and land three hits caused the antler to regrow, props to change, and pellet counts or trajectories to fail. The sequence was split into short causal segments. Generation handled character motion, while deterministic editing fixed count, trajectory, and contact.

## Final-gag measurements

- TAKO remains upright: `tako_crouches=false`
- Deer backs toward TAKO: 40.00–44.00 seconds
- Three face contacts: 44.33, 44.73, and 45.13 seconds
- Count: 3 pellets
- Cut: about 0.15 seconds after the third hit

When a generative model cannot lock down strict physics, counts, or collision points, moving those requirements to deterministic editing is reusable for other comedy and object-impact scenes.