video-workbench/investigations/2026-06-24-s14-reference-ca.../reviews/comparison-and-user-review.md

2.6 KiB

S14 Reference Comparison And User Review

Status

  • Document status: user decision recorded.
  • Codex local comparison: complete.
  • User decision: recorded in summary-and-rule-implications.md.

Outputs For Review

Arm Output Reference
A ../outputs/s14-anchor-board-reference-v1.png master character anchor board
B ../outputs/s14-side-crop-reference-v1.png side identity crop

Codex Comparison

Dimension Arm A: master board Arm B: side crop
Reference evidence Pass via visible image context Pass via visible image context
Identity preservation Strong overall identity and outfit preservation Stronger side-face identity preservation
Glasses / hair / beard Preserved Preserved more tightly in side profile
Coat / backpack Very strong Strong
S-14 half-profile function Good, but a little more three-quarter Better match
Environment-led MV feel Slightly better scene balance Slightly more figure-dominant
Main risk Less precise side-face control Portrait-size pull from crop

Codex Ranking

For S-14 specifically:

  1. Preferred primary reference: side identity crop.
  2. Useful supporting reference: master character anchor board, especially if the tool path supports multiple reference images.

If only one reference can be used for S-14, Codex recommends the side identity crop.

If multiple references can be used, Codex recommends:

Primary: character-anchor-v2-side-reference-crop-tight.png
Secondary: character-anchor-v2-codex-posture-test.png

The side crop should control face / glasses / hair / beard / age. The master board should control coat / backpack / posture / overall character continuity.

User Decision

The user concluded:

  1. The successful channel is gpt-image-2 Advisor prompt generation plus loading the real reference image into context before host image_gen.
  2. This successful channel should be used now and in future; API/Garden generation should not be the default path.
  3. Both the master anchor board and crop reference preserved strong character consistency.
  4. Near shots should prioritize crop references; distant shots should prioritize the anchor board.
  5. For 在路上, because most shots are landscape-led, aerial, or distant and only need character consistency, crops derived from the accepted anchor board are sufficient.
  6. For future projects with high identity requirements and many close shots, higher-resolution crop references may be generated through image-to-image from the accepted anchor board before production.