| name | refine-illustrations-iteratively |
| description | Refine AI-generated illustrations through successive, controlled visual adjustments while preserving approved details and character continuity. Use for iterative image-generation or image-editing sessions involving follow-up directions such as changing composition, viewpoint, aspect ratio, pose, anatomy, wardrobe, props, facial visibility, typography, atmosphere, or removing unwanted artifacts from the latest image. |
Refine Illustrations Iteratively
Build the final image as a sequence of controlled deltas. Treat the latest accepted image as the source of truth and preserve everything the user did not ask to change.
Maintain a visual state
Track three compact sets after every turn:
- Locked: approved elements that must survive future edits.
- Delta: the specific variables requested in the current turn.
- Watch: details likely to regress when applying the delta.
Infer locks from approval language such as "perfect," "great," "this is good," or a request that changes only one detail. Do not ask the user to restate established attributes.
Apply the refinement loop
- Anchor the scene. Establish subject, setting, mood, medium, and central visual idea.
- Solve structure. Adjust viewpoint, framing, aspect ratio, camera distance, placement, motion, geometry, and major pose before fine styling.
- Stabilize identity. Preserve the number of figures and their distinguishing traits, silhouette, proportions, hair, headwear, face visibility, and relationships to one another.
- Tune styling and props. Refine clothing, fabric, accessories, vehicles, devices, colors, and material details without redesigning locked elements.
- Add atmosphere and typography late. Introduce lighting, texture, glow, ornament, and exact text after the composition is stable. Preserve exact spelling, capitalization, line breaks, placement, and treatment.
