The Ninety-Second Launch Window
A regional bank needs three hero images for a mobile-app onboarding carousel. The brand team supplies a 40-word style brief: warm natural light, soft depth of field, diverse customers aged 25-45, no visible competitor logos, muted teal-and-sand palette. The art director runs a first batch of 40 text-to-image generations. Results: 31 images contain plausible but wrong text on signage, 9 have hands with six fingers, and every image drifts toward saturated orange rather than muted teal. The carousel ships in 90 minutes and only 3 images are needed. The art director has no time to retrain a model and cannot license stock in time.
1. The art director wants to re-run all 40 prompts with the single addition 'muted teal and sand palette' appended. Evaluate this plan and propose a better iteration strategy given the 90-minute constraint.
Read the worked answer
Appending a palette phrase to all 40 prompts is a weak fix for three reasons. First, it treats one symptom (color drift) while leaving the signage-text and hand defects untouched - those are prompt-structure problems, not style problems. Second, re-running 40 prompts burns most of the 90-minute budget on volume rather than on the 3 images actually needed. Third, a single appended clause competes with the existing 40-word brief; long competing descriptors dilute each other.
A better strategy:
- Triage by defect, not by batch. Discard the 31 signage images outright - text artifacts are the hardest to fix in post. Keep the 9 hand-defect images only if cropping removes the hands.
- Rewrite prompts structurally. Move style terms to the front, keep the brief under ~25 words, and add explicit negatives: 'no text, no signage, no logos.' For hands, specify 'hands out of frame' or 'hands relaxed at sides' rather than hoping.
- Generate in small targeted batches - 6-8 images per variant, changing one variable at a time (palette phrase, hand handling, composition). Stop as soon as 3 pass.
- Reserve 20 minutes for light generative editing (inpainting signage areas, color correction) rather than chasing a perfect generation.
The judgment call: under a hard deadline, reduce scope and fix defects by category, don't re-roll everything.
2. Two of the three finalists show a customer wearing a lanyard with unreadable but logo-like shapes. The bank's legal team flags 'no visible competitor logos.' What is the correct action and why?
Read the worked answer
Do not ship either image as-is. Even unreadable, logo-like shapes read as brand marks and can resemble real companies; 'plausible but wrong' marks are exactly the failure mode the brief forbids. Two acceptable paths:
- Regenerate with an explicit negative ('no lanyards, no badges, no logos, no text') and inspect at 100% zoom.
- Generatively edit the lanyard region (inpaint/erase) so the area is clean fabric, then re-inspect.
The reasoning: the constraint is about appearance of branding, not legibility. A human reviewer glancing at a carousel will perceive a badge shape as a logo. Legal review is a hard gate, not a style preference, so the cost of one more generation cycle is trivial against the risk of shipping an implied third-party mark.