# Technical information

This record concerns the Aha and spot-the-difference video production method introduced in Episode 6. Model and tool names are production-time records, not current recommendations.

## Finished specification for one work

- Challenge video: 22 seconds
- Answer version: 26 seconds
- Video: 1280×720, 24 fps, H.264
- Challenge timing: two-second opening hold, 18-second gradual change, two-second closing hold
- Answer: four-second before-and-after comparison
- Change curve: smoothstep
- Audio: none

## Technology by stage

| Stage | API, model, or tool | Purpose in this production |
|---|---|---|
| Requirements and candidate design | OpenAI Codex | Organized the one-sentence request, earlier works, and human evaluation into subjects and change proposals |
| Image production | OpenAI Image Generation | Created new reference images and reference-based edited after images. The model ID was not recorded, so none is guessed here |
| Image processing | Python, Pillow, NumPy | Standardized dimensions, limited the change region, composited before and after states, and produced answer images and contact sheets |
| Video rendering | Python, FFmpeg | Deterministically produced the 2 + 18 + 2-second challenge and appended the four-second answer comparison |
| Inspection | FFmpeg, ffprobe, Pillow, NumPy | Checked complete decoding, duration, resolution, frame rate, absence of audio, intermediate states, and locked regions |
| Evaluation tracking | JSON, Markdown, and human review | Kept OK, NG, idea duplication, and explicitly stated reasons separate and carried confirmed patterns into later trials |

Time interpolation, locking unchanged regions, answer screens, video rendering, and decode checks did not use generative AI. They used deterministic local processing that can reproduce the same output from the same inputs.

## Creating the gradual change

1. Prepare before-and-after images with the same dimensions and composition.
2. Limit the permitted change area as internal production data.
3. Lock every other pixel to the before image throughout the video.
4. Use smoothstep over 18 seconds so the change does not start or stop abruptly.
5. Add two-second holds before and after the change.
6. Create a four-second answer screen comparing the two endpoints and identifying the changed area.

This internal process must not be described as source material that the user supplied. The user states the desired subject and reviews the completed trials.

## Design patterns that worked better

- Keep the large main subject stable and change a less-attended surrounding detail.
- Limit each work to one meaningful change.
- Use repeated patterns, reflections, thin lines, or small groups that can change without breaking the meaning of the scene.
- Preserve shape and position while changing color, brightness, pattern, or local detail density.
- Keep enough difference for the answer to be understandable afterward.
- Describe how each idea differs from earlier work in subject, location, endpoint difference, and viewing experience.

## Failures observed in production

- Moving a large outline, such as an ear, or replacing a large object produced a double image at the midpoint.
- Broad change regions allowed unintended background or fur changes.
- Large movement or object replacement made some challenges immediately obvious.
- Very small changes remained unclear even after the answer.
- Technically valid work could still be rejected for naturalness, interest, or overlap with an earlier idea.

## Quality gates

### Before image production

- Does the idea duplicate one of the accepted works?
- Do the before and after images match in dimensions, composition, and subject position?
- Does the proposal require large outline movement or occlusion that simple interpolation cannot handle?
- Can the answer be explained in one sentence?

### Before video rendering

- Is there any difference outside the intended area?
- Are there leftovers, boundaries, or unrelated changes in the final state?
- Is there a double image or unnatural transparency at 50%?

### After video rendering

- Are the durations 22 and 26 seconds, with 1280×720 video at 24 fps?
- Does the entire file decode successfully?
- Do eight extracted frames and five progress stages show only the intended change?
- Is the answer not discovered too early?
- Is the difference clear after the answer?
- Are the human decision and stated reason recorded?

## Production and evaluation record

- More than 50 trials produced in half a day
- 57 evaluated works
- 32 accepted
- 22 rejected
- Three excluded for idea duplication

The project initially accepted only two or three videos out of ten. The finished making-of explains that after roughly ten rounds, eight out of ten were fully accepted and the remaining two were usable with minor adjustments.

The batches did not use identical conditions, so the figures are not described as a controlled performance comparison. They document a production process in which human feedback changed later trial criteria.

## Wording caution

These videos use the fact that gradual visual changes can be difficult to notice. The page must not claim that this is identical to the psychological Aha experience or that viewing the videos improves cognition or prevents medical conditions.