# Retention Studio Wan2.2 I2V asset workflow

Use this for motion-led short-form projects where local AI-video clips are only one layer of a deterministic Retention Studio.

## Production principle

Do not optimize for “AI video generated” as the win condition. Optimize for:

```text
hook clarity -> subject clarity -> controlled motion -> mechanism payoff -> renderer-controlled truth
```

AI-video assets provide photorealistic texture and motion. The renderer provides captions, labels, numbers, arrows, x-ray layers, risk meters, and any factual claim.

## Preferred pipeline

```text
mechanism explainer
-> mechanism shot plan
-> render manifest
-> asset plan
-> I2V prompt plan
-> keyframe generation/selection
-> native Wan2.2 I2V variants
-> clip review
-> deterministic final assembly
```

## Style direction for Hidden Money & AI Systems

Recommended look:

```text
photorealistic macro investigation + clean mechanism overlays
```

Examples:

- red used compact car in bright inspection bay
- hand hovering over finance/app confirmation
- invoice/document stack without readable details
- phone/feed close-up with no generated UI text
- desk/AI-chat scene with screen text blurred or replaced later

Palette:

- graphite/dark navy base
- white kinetic captions
- electric cyan for system/data flows
- amber for hidden cost/risk
- red only for twist/danger reveal

## Prompt-plan and job-package requirements

For each AI-video scene, produce:

- `keyframe_prompt`
- `i2v_motion_prompt`
- `negative_prompt`
- `wan_preset`
- clip requirements / review criteria

Then package the production handoff explicitly:

```text
I2V prompt plan
-> keyframe request package for ChatGPT Pro / image generator
-> seeded I2V clip jobs
-> ComfyUI run report
-> clip review package
```

Keyframe request packages should include target file paths, required dimensions, prompt, negative requirements, and an acceptance checklist. The checklist must reject readable text/logos/license plates, dark cinematic mush, weak subject silhouettes, and missing caption-safe space. For non-AI scenes, produce renderer-only notes instead of forcing Wan to draw diagrams.

Execution runners should be conservative and auditable:

- support `--dry-run` so missing keyframes and output paths are visible before GPU work starts
- skip jobs with `missing_keyframe` rather than inventing or falling back to T2V
- skip existing outputs unless the user explicitly asks to regenerate
- build native Wan2.2 workflow JSON from each job spec, not hand-edited throwaway workflow files
- write a run report with `total/succeeded/skipped/failed`, prompt IDs, output paths, and skip/failure reasons

## FPS policy

Do not treat low-fps test settings as final delivery settings. On a 16GB A5000, use conservative generation settings to prove subject/prompt quality, but deliver Shorts smoothly:

```text
preview/smoke: 49 frames @ 10 fps ~= 4.9s
better local production baseline: 61 frames @ 12 fps ~= 5.08s
final delivery: 24 fps after renderer/FFmpeg optical-flow interpolation or equivalent retiming
```

10 fps is acceptable only for technical smoke tests; it can visibly stutter on hopping, camera pushes, parallax, and other retention-critical motion. Prefer 12 fps generation for candidate clips, then interpolate/composite to 24 fps in the final renderer.

## Review criteria

Reject clips when:

- subject/action is unclear in first 0.5s
- output is dark, flickering, blurry, or abstract
- generated text/logos/plates/UI are prominent
- motion contradicts the retention beat
- there is no clean caption/overlay space

Do not call a clip “usable” just because ComfyUI succeeded. Technical success and content quality are separate gates.