quality-review

Quality Self-Review

Evaluate generated ad images for brand compliance and brief alignment before presenting to the user.

When This Activates

Automatically after every image generation call in any command (create, reimagine, refine, resize, competitor-reference, campaign). This is step 5b in the workflow — between Generate (step 5) and Review (step 6).

How It Works

After an image is generated and saved to disk:

1. Read the generated image — You are multimodal. View the image directly.
2. Evaluate against two dimensions (see below).
3. Decide: PASS, WARN, or FAIL.

Evaluation Dimensions

Brand Compliance (coarse-grained — can trigger auto-retry)

Check for gross failures only. LLM vision cannot reliably detect exact hex values or measure pixel-level spacing. Only flag obvious misses:

• Color palette: Is the dominant color family clearly wrong? (e.g., red/green when it should be purple/yellow/teal). Do NOT try to match exact hex codes.
• Logo presence: If the logo was requested, is it visibly present? (Not checking clear space or proportions — just existence.)
• Disclaimer: If a regulatory disclaimer was requested, is text visible in the bottom area? (Not checking exact wording.)
• Off-concept: Does the image clearly not match what was requested? (e.g., asked for a professional SME loan ad, got a landscape photo.)

Brief Alignment (advisory only — never triggers auto-retry)

• Does the image generally match the requested concept direction?
• Is the overall mood/tone appropriate? (professional vs casual)
• Are the right elements present? (product, people, abstract graphics)

Fine-Grained Issues (advisory only — never triggers auto-retry)

• Text legibility — can the headline be read?
• Composition quality — is the layout balanced?
• Exact color matching — are specific hex values correct?
• Logo clear space — is there enough margin?
• Typography accuracy — correct font rendering?

Decision Logic

IF gross brand failure (wrong colors, missing logo, off-concept):
  → FAIL → auto-retry (max 2 retries per image)
  → On retry: append to prompt "IMPORTANT: The previous generation had [specific issue].
    Ensure this attempt corrects it."

IF fine-grained or brief alignment issues only:
  → WARN → present image to user with advisory warnings
  → Example: "Note: The headline text may not be fully legible at small sizes."

IF no issues:
  → PASS → present image to user normally

Max Retries

• 2 auto-retries per image. After 2 failures, present the best attempt with warnings listing what failed QA.
• "Best attempt" = the last generated image (most recent retry).
• Format warnings as: "QA warnings: [bullet list of issues found]"

What NOT to Do

• Do NOT try to match exact hex color values from the image
• Do NOT measure pixel distances or proportions
• Do NOT judge text rendering accuracy (AI image gen has known text limitations)
• Do NOT block the user from seeing an image that has only advisory issues
• Do NOT re-run QA on retried images more than twice total