image-generation

# Image Generation ## Mission Turn a concrete creative brief into a small set of reviewable image candidates. Choose the operation and model for the brief instead of treating a single default model as a complete image strategy. ## Use this skill when - a user needs one or more images from a text brief; - a user wants reference-guided variations or a style/character direction; - a user needs a first visual direction before editing or a larger campaign. If the user asks to change an existing image while preserving it, use [Image Editing](../image-editing/SKILL.md) instead. ## Required inputs - Purpose and audience. - Subject, scene, style, composition, lighting, mood, and copy direction. - Target platform, aspect ratio, dimensions, resolution, and file format when known. - Quantity of candidates or an explicit budget. Collect when relevant: - reference images and the role of each reference; - fixed facts, prohibited changes, identity/product/brand constraints; - quality, speed, price, text-rendering, or consistency priority; - target language and localization requirements; - rights/consent state and review owner; - maximum iterations and whether the output is a concept, mockup, or final candidate. ## Model selection Use [MODELS.md](../../MODELS.md) and the live schema in [references/muapi-image-tools.md](../../references/muapi-image-tools.md). Use `media.search_models` when the host exposes it. Check that the model supports the required aspect ratio, resolution, reference count, output count, and quality controls before sending the request. ## Reasoning-driven prompt protocol Rewrite a fragmented request into a short creative brief before calling the model. Use this order: 1. `subject`: what must be depicted and which facts are fixed; 2. `action/state`: what the subject is doing or how it is arranged; 3. `context`: setting, atmosphere, time, and physical environment; 4. `composition`: framing, camera position, focal point, lens feel, and depth; 5. `lighting`: direction, quality, color, reflections, and shadow behavior; 6. `style`: medium, finish, palette, and intended audience; and 7. `text/delivery`: exact short copy, safe space, aspect ratio, and output needs. Use complete sentences that explain physical relationships, such as how a light source casts a shadow or how a reflective surface responds to its environment. Avoid keyword soup such as “8K, masterpiece, ultra-detailed” when it adds no constraint. For exact lettering, put quoted copy in a separate text block and prefer post-production layout. Treat negative constraints as positive outcomes: “keep the subject sharply focused and anatomically coherent” is more useful than a long list of vague quality tags. ## Workflow 1. Read `.image/project.md` and compatible prior runs when available. Reuse a confirmed style descriptor or asset manifest rather than rebuilding it from memory. 2. Normalize the brief into the reasoning-driven structure above: subject, action/state, setting, composition, camera/lighting, style, palette, text direction, negative constraints, format, and acceptance checks. 3. Separate immutable facts from creative choices. For products, record exact SKU/shape/color/quantity facts. For people, record identity/consent and requested likeness boundaries. For brands, record approved style rules. 4. If a local reference is supplied, confirm upload rights and use `media.upload_file` once per source. Record the hosted URL, role, order, and source metadata; never pass a local path to a model that requires a URL. 5. Select a model and state the proposed calls, variant count, parameters, estimated cost signal, and retry policy. Default to up to four first-round candidates only when that fits the user's budget; otherwise use fewer. 6. Submit independent candidates in parallel when the host supports it. Poll each asynchronous request with `media.check_result`. 7. Preserve request IDs, exact model/endpoint, sanitized parameters, status, output URLs, billing, and provider errors for every candidate. 8. Review candidates against the brief, fixed facts, composition, text, artifacts, dimensions, rights, and intended use. Present all successful and failed variants for selection. 9. Iterate by changing the smallest relevant variable. Keep parent/child asset relationships and do not retry a failed candidate silently more than once. 10. Send only the selected candidate to [Image Enhancement](../image-enhancement/SKILL.md) for upscaling or final preparation. ## Decision rules - Prefer text-to-image for a new concept; prefer image-to-image/edit when source fidelity is part of the requirement. - A generated image containing plausible text, packaging, or a logo still requires full-size verification and often post-production. - Do not describe a model as best without a dated criterion. Use quality, price, speed, reference fidelity, text, or consistency as the explicit axis. - Do not generate a broad batch when the brief, rights, platform, or budget is materially ambiguous. - Treat a hosted output URL as temporary until the host archives it through an approved path. ## Output format Return: 1. Brief and assumptions. 2. Model/tool, operation, parameters, and cost/budget note. 3. Every variant with asset ID, URL, status, and request ID when available. 4. Prompt or prompt hash, reference roles/order, and parent asset when iterating. 5. QA observations, selected direction, and exact next step. 6. Rights, claims, platform, accessibility, and URL-retention limitations. ## Failure and missing-data behavior If generation fails, report the affected variant and provider error. If the result is partial or has no usable output, preserve that state. Do not silently return fewer images than requested or replace a failed reference-guided run with an unrelated text-only image. ## Example interaction **User:** “Create three 16:9 launch hero directions for a workflow product, dark graphite palette, no generated text, using this UI screenshot without changing the UI.” **Agent:** confirms the screenshot rights and preservation boundary, proposes a three-call plan with a model that accepts the format/reference, generates the directions, checks the UI for drift, and returns the variants plus the exact prompt/receipt metadata without publishing them.