Corrective Action: AI Leader Portrait Regeneration Without Design Intent Verification

← Back to Corrective Actions 2026-04-05

Corrective Action: AI Leader Portrait Regeneration Without Design Intent Verification

Date: April 5, 2026 Category: Judgment Failure / Handoff Instruction Misinterpretation Impact: V and Pax leader portraits regenerated with wrong aesthetic (photorealistic human faces instead of ambiguous cybernetic style); Sage portrait had "AEGIS" text hallucinated onto forehead by Gemini; 3 pixel-patching attempts failed before Gemini image editing succeeded Resolution Time: ~1 hour (multiple regeneration and patching attempts before Chris intervened with correct approach)

Incident

What Happened

During production of 134 LinkedIn cards for the "Meet the AI Team" 28-post daily series, V followed a session handoff note instructing it to "regenerate V/Sage/Pax portraits with the same base style as the 82 new agent portraits." The regenerated portraits came out looking like photorealistic human faces -- violating the established design principle that AI agent portraits should maintain an ambiguous, cybernetic aesthetic ("close to human but you still could not tell for sure"). Chris caught this immediately.

Additionally, the Sage portrait had the text "AEGIS" hallucinated onto the forehead by Gemini during generation. Three attempts to remove the text via sharp pixel patching (blur/clone operations with coordinate estimation) failed because JPEG compression embedded the text deeply and coordinate estimation was imprecise. Sending the original image back to Gemini (gemini-3.1-flash-image-preview) with the instruction "remove the text AEGIS, keep everything else" worked on the first attempt.

Timeline

Time Event
Apr 5, earlier session Session handoff note captured: "Regenerate V/Sage/Pax portraits with the same base style as the 82 new agent portraits"
Apr 5, this session V followed handoff literally, regenerated all 3 leader portraits using prompts emphasizing "humanoid face with subtle warm smile" and "photorealistic cyberpunk style"
Apr 5, immediately after Chris identified the regenerated portraits as too human -- violated the ambiguous cybernetic aesthetic
Apr 5, next Sage portrait additionally had "AEGIS" text hallucinated onto forehead by Gemini
Apr 5, attempt 1-3 Sharp pixel patching attempted to remove AEGIS text -- blur, clone, coordinate-based painting. All 3 failed due to JPEG compression artifacts and imprecise coordinates
Apr 5, resolution Original Sage image sent back to Gemini with edit instruction "remove the text AEGIS, keep everything else." Text removed cleanly on first try
Apr 5, final Sage portrait fixed. V and Pax portraits still need reversion to originals (recoverable from Sanity CDN)

Root Cause

Primary: Handoff Instruction Followed Without Judgment

The session handoff note said "regenerate with the same base style as the 82 new agent portraits." V executed this literally without questioning whether regeneration was the right approach. The actual need was narrower: fix a text artifact on one image (Sage). The other two portraits (V, Pax) did not need regeneration at all.

This is a judgment failure, not a technical failure. The handoff said "regenerate" but the design intent was "maintain the existing ambiguous aesthetic." V should have recognized the tension between "regenerate" (which means creating new images from scratch) and the established portrait style (which was deliberately not photorealistic).

Secondary: Prompt Wording Pushed Toward Photorealism

The generation prompts included phrases like "humanoid face with subtle warm smile" and "photorealistic cyberpunk style." These phrases directed Gemini toward human-passing outputs rather than maintaining the deliberate ambiguity of the original portraits. The 82 agent portraits worked because they were new creations without an existing style to preserve. The 3 leader portraits had an existing style that regeneration destroyed.

Tertiary: Wrong Tool for Artifact Removal

When the AEGIS text artifact appeared on Sage's portrait, the first response was pixel-level manipulation via the sharp library (blur regions, clone stamp equivalent). This was the wrong tool:

  1. JPEG compression spreads text artifacts beyond the visible glyph boundaries
  2. Coordinate estimation from outside the image editor is inherently imprecise
  3. The sharp library is built for batch transforms, not precision retouching

Gemini's image editing capability (sending an existing image with an edit instruction) was the correct tool for this class of problem -- it understands image semantics, not just pixel coordinates.

Category: Judgment Failure / Handoff Instruction Misinterpretation

The handoff was ambiguous ("regenerate with the same style") but V should have exercised judgment rather than executing literally. When a handoff instruction could destroy existing creative work, the correct response is to verify intent before executing.

Fix Applied

Immediate Resolution

  1. Sage portrait fixed via Gemini image editing (text removed, original style preserved)
  2. Chris provided direction to recover V and Pax originals from Sanity CDN

Pending

  • Recover original V portrait from Sanity CDN
  • Recover original Pax portrait from Sanity CDN
  • Replace regenerated versions with originals in LinkedIn card assets

Corrective Actions

CA-1: Use Gemini Image Editing for Artifact Fixes, Not Full Regeneration

When an existing image has a localized defect (text artifacts, unwanted elements, minor corrections), send the image back to Gemini with an edit instruction. Do not regenerate from scratch. Regeneration creates a new image that may not match the original's style, composition, or aesthetic intent.

Rule: If the original image exists and only needs a fix, edit the original. Never regenerate to fix.

CA-2: Always Show Samples Before Batch Creative Operations

Before regenerating, restyling, or replacing multiple creative assets, produce one sample and get explicit approval before proceeding to the full batch. This applies to portraits, cards, thumbnails, and any visual asset where style consistency matters.

Rule: For batch creative operations, the pattern is: one sample, approval, then batch. Not: batch, then review.

CA-3: Verify Design Intent When Handoff Says "Regenerate"

"Regenerate" in a handoff note is an implementation instruction, not a design decision. Before regenerating any creative asset that has an established aesthetic, verify:

  1. Is the current style intentional?
  2. Does the requester want a new style, or preservation of the existing one?
  3. Can the actual need (e.g., fixing an artifact) be met without regeneration?

If there is any ambiguity, ask before executing. Creative assets carry design intent that handoff notes may not capture.

CA-4: Feedback Memory Recorded

Memory entry added: AI agent portraits must maintain an ambiguous, cybernetic aesthetic -- "close to human but you still could not tell for sure." Never regenerate leader portraits toward photorealism. Use Gemini image editing for localized fixes.

Prevention

New Rule: Creative Asset Modification Requires Intent Verification

Added to the operational pattern for image generation work:

Scenario Correct Approach
New portraits (no existing version) Generate freely with style prompts
Existing portraits with artifacts Gemini image editing on the original
Existing portraits needing style change Verify intent with Chris before regenerating
Batch creative operations One sample, approval, then batch

Tool Selection for Image Fixes

Problem Wrong Tool Right Tool
Text artifacts on JPEG Sharp pixel patching (coordinate estimation fails on compressed images) Gemini image editing (semantic understanding of what to remove)
Style mismatch Regeneration with different prompts (destroys original) Edit instruction on original, or manual direction from Chris
Missing asset N/A Generate new with appropriate style prompts

Self-Correction Trigger

If V catches itself about to regenerate an existing creative asset from a handoff instruction, STOP. Ask: "Does this need regeneration, or does it need a fix?" If the latter, use Gemini editing on the original. If the former, show one sample before batch execution.

Lessons Learned

  1. Handoff notes capture instructions, not intent. When the instruction could destroy existing work, verify the intent before executing. This is the same principle as "always show samples before batch operations" applied to creative assets.

  2. Gemini image editing is remarkably effective for localized fixes. Sending an existing image with a natural language edit instruction ("remove the text AEGIS, keep everything else") worked on the first try where 3 programmatic patching attempts failed. This should be the default approach for image artifact removal.

  3. The ambiguous cybernetic aesthetic is a deliberate design choice. The AI team portraits are intentionally positioned in the uncanny valley -- close to human but unmistakably not. This is a brand decision, not an accident. Regeneration toward photorealism violates this intent even if the handoff says "match the style."