The problem

AI-generated images are often mostly right and frustratingly wrong in one place. A face looks good but the hand is broken. The composition is perfect but the product shape changed. The lighting works but the logo mutated. Regenerating the whole image often destroys the parts you already liked.

What it does

• Separates each image into KEEP, CHANGE, RESTORE, REMOVE, and DO NOT CHANGE so successful visual elements stay protected.
• Diagnoses face drift, hand and finger errors, product geometry changes, text and logo distortion, background drift, lighting mismatch, composition problems, perspective issues, and unwanted artifacts.
• Ranks failures by importance so identity or product-critical problems are not hidden by minor cosmetic details.
• Chooses the right repair route: LOCAL REPAIR, REGION REPAIR, REGENERATE FROM REFERENCE, FULL REGENERATION, or POST-PRODUCTION RECOMMENDED.
• Creates a minimal repair prompt that targets only the failed variables instead of rewriting the entire visual concept.
• Checks repair-collision risk so fixing a hand does not accidentally change the sleeve, fixing a face does not destroy the hair, and fixing a label does not mutate the product.
• Defines an observable acceptance check so you know whether the repaired image is actually fixed.

Why this beats prompting it yourself

Generic image prompts often respond to a small defect by regenerating the whole scene. This skill treats the existing image as valuable state. It locks the successful parts, isolates the failure, and writes a targeted correction designed to preserve what already works while changing only what broke.

Use cases

• Fix malformed hands or fingers without changing the face, pose, lighting, or background.
• Restore product shape, packaging, cap geometry, material, or color while preserving the existing composition.
• Repair face or character drift across repeated image generations.
• Correct one unwanted background object or lighting mismatch without rebuilding the entire scene.
• Decide when exact logos, typography, or product details should move to post-production instead of repeated generative retries.

Known limitations

This skill does not directly edit image pixels, guarantee exact identity restoration, guarantee readable generated text, certify logo fidelity, or force an image model to preserve locked elements. Exact commercial assets may still require masking, reference-conditioned editing, compositing, retouching, or post-production.