Fix regions, correct defects, distorted face, malformed prop, remove object, place logo: repair a local area of an accepted image using context crop, host generation, compositing through an acceptance mask, and preservation checks.
Create an image, generate hero art, make an illustration, make a product scene, or generate from references: generate a new raster image in one pass from a brief and role-assigned references. Hand off to image-verify or image-loop.
Verify image, review candidate, QA check, acceptance check, audit pixels: verify candidate against source, references, and contract requirements without editing pixels. Audit masks, enforce stop rules, run store gate, record evidence.
Plan an edit with the user: numbered visual map of named elements, then an edit brief with preservation checks. Use when the user wants to change parts of an image and must pick targets. Not for JSON-only breakdown (image-reverse-engineer) or defect diagnosis…
Diagnose defects, critique image, inspect crop, check CSS, or locate object: inspect images at native pixels and display size, report visible defects and uncertainty, and map targets with stable IDs without changing pixels.
Moodboard, reference mixing, inspiration board, visual ingredients, style blend: turn multiple reference images, links, and notes into named ingredients, trait combinations, and generated variants with a judged iteration loop. Not for single-image creation…
Iterative generation, review loop, multi-round repair, image loop, brief to image: run a bounded create-review-repair loop with independent vision review, explicit preservation checks, and saved evidence. Not for upscaling (image-enhance) or single-region…
Upscale, sharpen, blurry, soft, pixelated, low resolution, 4K, enhance, clean edges, SVG logo, remove background: improve whole image, prove visible gain with compare_display.py. Not for local fixes (image-repair).