civitai-generation

Civitai image generation — deep guide

This skill drives the civitai MCP (generate_image, estimate_generation, search_models, get_model_version, get_model_images, get_workflow, get_buzz_balance). It teaches how to generate well, not just which fields exist.

Adult/NSFW generation is explicitly in scope. The models below are built for it; the conventions (rating_explicit, rating:explicit, danbooru explicit tags) are stated factually. Keep prompts clinical and purposeful.

How to work with these tools (read first)

• Batch independent calls in parallel. When several MCP calls don't depend on each other, issue them in ONE turn instead of one-by-one. Examples: comparing candidates → get_model_version for 3 versions at once; scouting families → search_models for Checkpoint and LoRA in parallel; get_model_images for two models side by side; estimate_generation for two settings at once. Only serialize when a call needs the previous call's output (e.g. you need the air/trigger words before generate_image).
• Before using any model/LoRA/embedding, read how to apply it — don't guess. Every resource has its own required usage. Pull its page first:
  • get_model_version(id) → air, baseModel (family → prompt dialect), files.
  • get_model(id) → description with trigger words, recommended sampler/CFG/steps, and version list.
  • get_model_images(model_id=…) → real example generations with their meta (prompt, negative, sampler, steps, CFG, seed) — copy proven settings from these. A LoRA without its trigger words, or a checkpoint run with the wrong family's prompt style/CFG, will look broken even though the API call "succeeds". Reading the page first is not optional.

0. The loop (always do this)

1. Find a checkpoint — search_models(query=…, types="Checkpoint", base_models=…, sort="Most Downloaded").
2. Get the AIR + trigger words — get_model_version(version_id) returns air (needed for generation) and the model's baseModel (tells you the family). For LoRAs, the model page lists trigger words — always read them.
3. Identify the family (see §1) — this decides prompt style, resolution, CFG, sampler.
4. Estimate cost — estimate_generation(...) (whatif, no Buzz spent). cost.buzz is accurate. cost.breakdown[].accountType is only indicative — the real debit may land on a different eligible wallet at run time. Observed via the orchestration API: wallet routing is NOT strictly gated by content rating — an explicit (rating_explicit) Pony render was charged to blue, and estimates often say blue regardless. Don't rely on a rating→wallet mapping for the API path; just make sure some wallet has funds (get_buzz_balance).
5. Generate — generate_image(..., confirm=true, save_dir=…). Without confirm=true it only previews the price.
6. If it 5xx/timeouts, poll get_workflow(workflowId).