bulk-image-generation-workflow

Bulk Image Generation Workflow

This skill orchestrates around one tool: mcp__ideogram__generate_images_bulk(prompts: list[str], ...). It accepts 1–500 prompt strings and submits them as one async background job; results render into a carousel as each image completes rather than blocking on the whole batch.

The constraint the whole skill is built on: every non-prompt parameter (aspect_ratio, resolution, rendering_speed, style_type, negative_prompt, seed, magic_prompt_option, custom_model_uri, collection_id, private) is shared across the entire batch — there is no per-prompt override. On-brief variation at scale has to come from differences in the prompt text itself, not from differences in call parameters. Note the caveat: style_type, negative_prompt, seed, and magic_prompt_option only take effect when custom_model_uri is set (the custom-model v3 path); on the default v4 path they're ignored, so passing them there does nothing.

This skill sits between collections-management's pure-orchestration pattern and the pure prompt-composition skills' pattern: mostly orchestration, with one prompt-composition component — turning one caption into N prompts — covered by references/variation-strategy.md.

Before you start

Read references/variation-strategy.md before drafting the prompt array (steps 2–3 below) and references/review-culling-guide.md before reviewing results (step 6 below). Both apply before you touch the workflow.

Workflow

1. Resolve the base caption

Use an existing locked structured JSON caption if the user or project context already has one. Otherwise build one quickly using skills/ideogram-prompt/references/json-caption-schema.md's schema and ideogram-prompt/SKILL.md's precise-mode guidance, rather than re-deriving the schema here. This caption's style_description object is what stays locked for the whole batch.

2. Choose the variation axis

Per references/variation-strategy.md, decide what changes per prompt — pose, prop, angle, framing, a swapped compositional_deconstruction element — while style_description stays byte-for-byte identical across every prompt in the batch. If the axis isn't obvious from the user's request, confirm it with them before drafting.