image-editing

# Image Editing ## Mission Make a controlled change to an existing image. Use this skill when the user says “edit,” “remove,” “replace,” “reframe,” “keep the same,” or supplies a source image whose identity, product geometry, or composition matters. ## Required inputs - Source image URL or local file. - Exact requested change. - Details that must remain unchanged. - Target aspect ratio, dimensions, file format, and quantity when relevant. Collect when relevant: - additional reference images and their roles/order; - mask or region guidance when the selected model supports it; - identity, product, logo, label, or packaging rights; - quality, speed, cost, and iteration priorities; - whether the result is a concept, mockup, or final commercial candidate. ## Model selection Use the editing/image-to-image entries in [MODELS.md](../../MODELS.md) and the live schema described in [the MuAPI tool reference](../../references/muapi-image-tools.md). Do not silently route an edit to text-to-image when the user expects source fidelity. ## Workflow 1. Inspect the source and create two lists: `change` and `preserve`. Include fixed geometry, identity, labels, logos, text, colors, quantity, lighting, and camera position when applicable. 2. Confirm rights to upload the source and any additional reference. Do not use a saved private identity or brand asset without an explicit match. 3. If the source is local, upload it through `media.upload_file` and record the hosted URL, role, order, and source metadata. Pass only hosted URLs accepted by the selected endpoint. 4. Build an edit prompt with the requested change first, followed by an explicit preservation block and unwanted-change constraints. Keep exact copy, labels, logos, and claims out of a regeneration request when approved source art can be composited instead. 5. Select the narrowest editing model that supports the source count, target ratio, and required operation. State the planned call, estimated cost signal, and retry policy before execution when unknown. 6. Generate a bounded preview. Poll with `media.check_result` and preserve the request ID, status, output URLs, billing, and provider errors. 7. Compare source and output for unintended changes. Check identity/product fidelity, geometry, text, edges, shadows, repeated objects, artifacts, crop, and requested region. 8. On feedback, change one relevant variable at a time. If the model cannot preserve a fixed detail, recommend compositing or a different supported operation rather than claiming success. ## Output format Return: - source asset and reference roles; - requested change and preservation block; - model/tool, parameters, and output dimensions; - output URL(s), asset IDs, request status, and receipt link when saved; - changed details, preserved details, observed drift, and required post-production; - rights, text, claims, and URL-retention limitations. ## Failure and missing-data behavior If the source cannot be uploaded, the model does not accept the requested reference, or the output is incomplete, preserve the failure and return the smallest supported alternative. Do not silently return a fresh image as though it were an edit. Retry a failed variant at most once without new user input.