3.6 KiB

Raw Blame History

Comparison

Rubric

Scores are subjective prompt-text scores from 0 to 5. They judge whether the prompt is likely to be controllable and reusable in video-workbench. They do not judge generated image quality.

Dimension	Meaning
D1 Intent Decomposition	Separates goal, subject, scene, and output use.
D2 Reference Role Control	Clarifies which reference preserves identity, wardrobe, pose, or style.
D3 Composition Specificity	Defines camera, framing, aspect ratio, environment ratio, and motion.
D4 Failure Prevention	Names predictable failure modes and avoids them.
D5 Workflow Fit	Fits `projects/<project-id>/...` style production records and later review.
D6 Model Usability	Gives image models concrete visual instructions without overloading readable text.

Scores

Case	Direct Total / 30	Advisor Total / 30	Main Difference
T01 Identity makeup still	20	27	Advisor treats the result as `identity_ref`, not a final cinematic scene.
T02 Wardrobe fullbody anchor	18	27	Advisor explicitly downgrades face likeness and prioritizes outfit geometry.
T03 MV side-walking shot	19	29	Advisor separates identity, wardrobe, pose, and style references and guards against portrait drift.
T04 Muddy boots detail	22	28	Advisor correctly avoids using identity as the primary reference.
T05 Dense explainer slide	18	27	Advisor prevents the common error of asking the image model to render final readable text.

Average:

Prompt Type	Average / 30
Direct	19.4
Advisor	27.6

Findings

At the prompt-text level, the Advisor workflow is materially stronger than direct prompting for this use case.

The main improvement is not prettier language. It is production control:

It states whether an output is an identity anchor, wardrobe anchor, final shot, detail shot, or slide background.
It splits references into roles instead of treating all references as the same kind of input.
It encodes known failures from the prior run, especially portrait drift and overusing face identity.
It creates prompts that can be versioned, reviewed, and revised per shot.

This supports the restructuring premise:

GPT should output director cards and acceptance standards.
Codex / video-workbench should render final image prompts using gpt-image-2 Advisor logic.

Limits

This does not prove final image quality. It proves only that the Advisor prompt artifacts are higher-quality production inputs under a text rubric.

The next validation level should be:

Pick 2-3 prompts from this investigation.
Generate images on the same image platform.
Judge image outcomes with the same acceptance criteria.
Record whether the Advisor prompts actually reduce failure rates.

Path Implications

The investigation supports adding a project-local character package folder for human-character projects:

projects/<project-id>/makeup-still/
  identity-prompt-v1.md
  identity-ref-v1.png
  wardrobe-prompt-v1.md
  wardrobe-ref-v1.png
  pose-ref-notes.md
  acceptance-notes.md

This folder should exist only when a project needs a recurring human / character identity system. Non-character projects should not carry the folder.

Stable Codex-side knowledge should not be buried in a single project. After owner acceptance, summarize this into docs/ as video-workbench operating knowledge, likely:

docs/image-prompt-advisor-workflow.md
docs/makeup-still-directory-policy.md
docs/reference-strategy-and-review-rubric.md

Do not write those stable docs before the owner accepts this investigation result.

3.6 KiB Raw Blame History