Generate images using the public /v2/inferences endpoint.
For the highest quality prefer RD Pro styles (rd_pro__*); they support reference_images for
character/style consistency, and most go as small as 12x12 px (check list_available_styles
for each style's limits) — a small target size is never a reason to switch to a cheaper
model family. Style ids are opaque strings with no uniform format (some RD Fast styles
appear as "default:rd_flux"); take them verbatim from the catalog and never infer
capabilities from an id's prefix. For animation styles prefer start_inference_job +
get_inference_job instead — animations are long-running, and a failed animation is worth
one retry with identical parameters (failures auto-refund).
Field-tested workflow rules: N distinct items = N individually usable images (separate
calls or num_images=N), never one sheet/grid image unless a sheet IS the deliverable.
Variants of ONE image (seasons, day/night, palettes) = generate the base once, then derive
each variant with the image_edit tool ("... keep the exact same composition") — independent
generations of the "same" scene come out unrelated. Converting an existing image INTO pixel
art is rd_pro__pixelate with input_image; reference_images-based generation re-imagines
rather than converts. To animate an image you already have, use rd_advanced_animation__*
with input_image (fixed-format rd_animation__* styles generate their own subject from the
prompt instead).
Use `input_image` for the main source image, `reference_images` for extra per-inference
guidance, and `style_reference_images` only on create_user_style/update_user_style.
The response excludes raw base64 image payloads to keep MCP outputs compact.