# Technical information

This record concerns the 127.77-second, 23-shot “Bitter-Melon Harvest and Deer Control” main film shown near the beginning of the full making-of video, not the entire making-of program. Model names are production-time records, not current recommendations.

## Finished specification

- Final version: v009
- Duration: 127.77 seconds (127.75-second design value)
- Shots: 23
- Video: 1280×720, 24 fps, H.264
- Audio: AAC, present
- Speech speed: 1.0× for all 20 lines

## Technology by stage

| Stage | API, model, or tool | Purpose in this film |
|---|---|---|
| Requirements and direction | OpenAI Codex | Organized the rough first request, later materials, previous-film callback, 23-shot cause and effect, scale, ending, and revisions |
| Video and spreadsheet review | Codex multimodal inspection | Inspected source motion over time, screenshots, the Excel storyboard, TAKO’s scale, fence overlays, and final frames |
| Excel storyboard | Python, openpyxl, Pillow | Managed 23 shots, representative images, dialogue, captions, action, sound, and review notes |
| Missing reference images | OpenAI Image Generation | Designed the harvest report, deer jump, and original colossal figure. The model ID was not recorded, so none is guessed here |
| Generated video | Google `gemini-omni-flash-preview` | Generated S002’s harvest report, S008’s deer jump, and S023’s approaching colossal figure |
| Final TAKO voice | Chatterbox TTS 0.1.7 / `ChatterboxMultilingualTTS` | Synthesized all 20 expressive lines from a roughly 20-second private voice reference using zero-shot voice synthesis |
| Audio post-processing | Python, SoundFile, NumPy | Normalized levels, controlled peaks, and inspected WAV files without changing speech speed |
| Live-action formatting | FFmpeg | Preserved vertical aspect ratio, created blurred side video, standardized 24 fps, extracted intervals, and retimed visuals to speech |
| Compositing, captions, and effects | FFmpeg, libx264, AAC, ASS | Deterministically composited TAKO, fence overlays, measurements, captions, ambience, sound effects, and the final sequence |
| Image processing | Pillow, OpenCV | Prepared transparent assets, locked scale, produced review images, and inspected frames |
| QA | FFmpeg, ffprobe, OpenCV, openpyxl, and human review | Checked full decode, duration, frame rate, audio, silence, Excel, midframes for all 23 shots, and critical motion |

## Fitting vertical live action into 16:9

The source videos were not stretched horizontally. The original moving pixels stayed in the center, while enlarged blurred copies of the same moving image filled the sides. The Excel storyboard showed representative frames, but the final film used the source MOV/MP4 motion, camera movement, vegetation, and ambience over time.

## Matching TAKO’s scale to the real field

- TAKO: about 160 cm tall including the straw hat
- Inner fence: 140 cm
- Gold tape: about 2 m above the ground

TAKO was placed on the same ground plane and at the same depth as the comparison object, then checked by pixel height. Because generated images caused position and scale to drift, S017, S022, and S023 reused the same transparent TAKO asset, coordinates, and scale in local compositing. For the final colossal-figure shot, generation handled the distant background action while TAKO stayed locked from the previous shot.

## Separating generation from exact elements

Fence lines, the 140 cm and 2 m measurements, captions, and the three-panel summary were fixed locally. The generator was not expected to preserve real fence geometry, exact text, measurements, and character scale at the same time. Generation handled motion and missing imagery; deterministic editing handled counts, text, dimensions, coordinates, and display timing.

## Synthetic speech based on the creator’s voice

Early versions used a different TTS system. The final version used Chatterbox TTS 0.1.7 after the creator compared restrained, natural, and expressive samples and selected the expressive profile.

Speech was generated first at natural speed, then shot duration followed this rule:

`shot duration = max(3 seconds, voice duration + 0.55 seconds)`

The result was rounded up to the next 0.25-second boundary. Speech remained at 1.0×. When five lines needed pronunciation corrections, only those lines were regenerated using separate spoken input; the other approved recordings were preserved. The private voice reference never entered the public deliverables.

## What the AI organized or proposed, and what a person checked

- Organized “deer cleared 140 cm, so tape was added at about 2 m” into a viewer-facing causal explanation
- Proposed a final three-panel review of the inner area, outer perimeter, and entrance
- Positioned TAKO as a generated on-location reporter
- Required human verification against the actual footage, measurements, and fence locations
- Omitted `シカ再び.mp4` when it weakened the central story
- Finalized the ending with the same TAKO remaining visible as an original colossal figure takes two steps before the cut to black

## Final QA

| Check | Result |
|---|---|
| Full decode | Completed without errors |
| Duration | 127.77 seconds measured; 127.75 seconds designed |
| Resolution and frame rate | 1280×720, 24 fps |
| Audio | AAC track present |
| Speech speed | 1.0× for all 20 lines |
| All shots | Midframes reviewed for all 23 shots |
| Critical action | Deer jump, scale, fence overlays, two colossal footsteps, and final blackout reviewed |
| Excel | Opened successfully; 24 sheets |