7.3 KiB

Raw Blame History

Character Transparent Slices Review V1

Status

Review status: not accepted by user.
Generation status: completed first five-slice batch.
Prompt/spec source: visual-system/characters/character-transparent-slices-prompt-v1.md.
Main identity source: accepted v2 character anchor board and character-identity-lock.md.
User rejection reason: the generated slices no longer look like the accepted v2 character anchor.
Next gated action: repair one slice at a time, starting from identity-critical front full body, before any style-anchor or five-shot work uses transparent character assets.

Generated Assets

Final alpha PNG slices:

Slice	Final file	Source chroma file
Front full body	`visual-system/characters/character-slice-front-v1.png`	`visual-system/characters/character-slice-front-v1-chroma-src.png`
Side full body	`visual-system/characters/character-slice-side-v1.png`	`visual-system/characters/character-slice-side-v1-chroma-src.png`
Back full body	`visual-system/characters/character-slice-back-v1.png`	`visual-system/characters/character-slice-back-v1-chroma-src.png`
Walking side	`visual-system/characters/character-slice-walk-side-v1.png`	`visual-system/characters/character-slice-walk-side-v1-chroma-src.png`
Walking back	`visual-system/characters/character-slice-walk-back-v1.png`	`visual-system/characters/character-slice-walk-back-v1-chroma-src.png`

Preview board:

visual-system/characters/character-transparent-slices-preview-v1.png

Execution Mode

gpt-image-2 Skill mode check:

{
  "mode": "B-or-C",
  "recommendation": "host-or-advisor",
  "garden_mode_enabled": false,
  "has_api_key": false,
  "base_url": "https://api.openai.com/v1",
  "model": "gpt-image-2",
  "env_flag_value": "(unset)"
}

Execution interpretation:

Garden local generation was not enabled.
Host image generation was used as Mode B.
Each slice was generated separately.
The built-in image tool does not expose a direct output-path argument; generated defaults were copied into the project.
The built-in image tool also does not expose native transparent-background control, so each asset was generated on a flat green chroma-key background and converted locally to alpha PNG.

Generation Source Mapping

Default host-generated source files were copied as follows:

Slice	Default source file
Front	`C:\Users\wangq\.codex\generated_images\019ef2f7-e630-7303-a0e8-e5e7a86912fc\ig_0cdcfc2a4d6685cc016a3a33d85a548191ace8cb7e4caa0623.png`
Side	`C:\Users\wangq\.codex\generated_images\019ef2f7-e630-7303-a0e8-e5e7a86912fc\ig_0cdcfc2a4d6685cc016a3a341539308191833e4f10d980865e.png`
Back	`C:\Users\wangq\.codex\generated_images\019ef2f7-e630-7303-a0e8-e5e7a86912fc\ig_0cdcfc2a4d6685cc016a3a34912710819196953af31c806bb2.png`
Walk side	`C:\Users\wangq\.codex\generated_images\019ef2f7-e630-7303-a0e8-e5e7a86912fc\ig_0cdcfc2a4d6685cc016a3a34c91b3c819182e18394180e4000.png`
Walk back	`C:\Users\wangq\.codex\generated_images\019ef2f7-e630-7303-a0e8-e5e7a86912fc\ig_0cdcfc2a4d6685cc016a3a3503e1d481918e8786806eb9a450.png`

Transparency Validation

Local alpha conversion used:

C:\Users\wangq\.codex\skills\.system\imagegen\scripts\remove_chroma_key.py

Python environment:

C:\Users\wangq\.cache\codex-runtimes\codex-primary-runtime\dependencies\python\python.exe

Validation result:

File	Mode	Size	Transparent pixels	Partial pixels	Corner alpha
`character-slice-front-v1.png`	RGBA	864x1821	959873	9220	`[0, 0, 0, 0]`
`character-slice-side-v1.png`	RGBA	864x1821	1041950	8430	`[0, 0, 0, 0]`
`character-slice-back-v1.png`	RGBA	864x1821	983848	8599	`[0, 0, 0, 0]`
`character-slice-walk-side-v1.png`	RGBA	1024x1536	1195863	7495	`[0, 0, 0, 0]`
`character-slice-walk-back-v1.png`	RGBA	864x1821	1044861	8060	`[0, 0, 0, 0]`

Validation interpretation:

All five final files have an alpha channel.
All four corners are fully transparent.
Partial alpha pixels exist around edges, which is expected from soft matte extraction.

Visual Review

Local preview file:

visual-system/characters/character-transparent-slices-preview-v1.png

Observed strengths:

All five figures are full body and fully inside the canvas.
The long dark coat, backpack, old boots, longer hair, glasses, and fuller beard system remains coherent enough for this first slice batch.
The walking side pose is upright and does not repeat the rejected v1 hunched walking issue.
The walking back pose reads as continuing forward rather than collapsed.
Back-view identity is usable: backpack, long coat, hair outline, and boot stride are clear.

Observed risks:

These were generated by host prompt-only Mode B. The local v2 anchor image and source photo were not attached as binary image inputs to the image tool.
Therefore, same-person fidelity is based on the approved textual identity lock plus generated-context continuity, not on a strict image-reference pipeline.
The front / side / back likeness is coherent at the character-system level, but should still be treated as a candidate set, not a final locked production asset set.
The side standing slice has a naturally forward profile and heavy coat; it should be checked by the user against the "not hunched" requirement.

Checklist

Check	Result	Notes
Five requested slices exist	Pass	Front, side, back, walk-side, walk-back generated.
Final files are alpha PNG	Pass	RGBA mode; transparent corners confirmed.
Full body inside canvas	Pass	No obvious crop in local preview.
Same character system	Pass with note	Coherent coat / backpack / boots / hair / beard system; not strict binary-reference identity.
Walking side posture	Pass	Upright chest and straighter spine; not the rejected v1 hunch.
Walking back posture	Pass	Continues forward without collapsed shoulders.
Back-view recognition	Pass	Backpack, coat, hair outline, and boots are stable.
Avoid fashion / hero drift	Pass	Practical road-worn silhouette, not fashion editorial.
Transparency edge quality	Pass with note	Soft alpha edge present; user should inspect at final compositing scale.

Recommendation

Do not use this slice set for production prompts.

Treat it as a failed identity-preservation test:

Keep the files as failed-trial evidence.
Do not delete them unless explicitly requested.
Do not promote this transparent-slice workflow into reusable rules.
Repair the slice workflow before generating more character execution assets.

Root cause assessment:

The host tool produced five separate prompt-only generations.
The accepted v2 anchor and source photo were recorded as references, but were not attached as strict binary image inputs.
Therefore the model regenerated a similar "road traveler type" instead of preserving the accepted v2 person.

Required process change:

Transparent character slices must be repaired one at a time.
The next attempt should be reference-bound or crop-derived from the accepted v2 board, not independent text-only generation.
The first repaired slice should be front, because face / hair / beard / glasses identity drift is easiest to judge there.

7.3 KiB Raw Blame History