Generate a cohesive SET of custom images with TRANSPARENT backgrounds, each one a separate isolated subject sharing one visual style (icons, logos, sprites, UI assets that need to drop onto any backdrop). The style parameter says how everything is drawn; the subjects parameter says what to draw. The style can also come from reference images via the styleReferences parameter - alone, or best combined with the style text (text plus references holds a style tightest), so an existing set can be extended in its original style across many calls. Returns a zip download URL. Each call costs 1 credit. When a solid colored background fits the user's use case, generate_image_set (1 credit per call) is the faster choice. Run generation calls sequentially, never in parallel - only one generation runs at a time per API key.

Output formats: PNG (lossless) or WebP. JPG is not supported because it has no alpha channel.

Size and quality considerations match generate_image_set: leaving width and height unset delivers native resolution (best quality, varies between generations); fixing them resamples to that box, and fixing one axis lets the other hug each subject.

IMPORTANT - style and subject description rules for best results: The style applies to every image in the set, so it is what keeps them visually consistent. Put HOW the images are drawn (technique, palette, surface treatment) in the style, and make each subject description only about WHAT that one subject is, not how it looks. A quick test for any phrase: is it WHAT the subject is, or HOW it is drawn? HOW belongs in the style, shared across the set.

• Get the style and the subject descriptions right with the user before you call. When their request puts how an image is drawn, a background, or a scene into a subject description (or names the subjects to draw in the style rather than as separate entries in the subjects list), fix it as you compose the call: routine moves of shared technique into the style you can just make, but when a change drops or alters something they explicitly asked for, tell them what you are adjusting and why first. Each call costs a credit, so it is worth getting this right up front rather than spending one on a framed or scene-filled result.

• The style describes the visual treatment of the images (e.g. 'watercolor', 'pixel art', 'stained glass'). It must NOT mention background color, image count, layout, or sizing.

• Do not list the subjects to draw in the style (e.g. 'illustrations of a fox, an owl, and a deer'); the style is only the shared visual treatment, and the subjects belong in the subjects list, one per entry. A category or theme word is fine (e.g. 'insect illustration').

• Do not put background color or background descriptions in the style or subject descriptions.

• Avoid framing the style as a type of painted canvas ('oil painting', 'acrylic painting', 'gouache painting', 'pastel painting'). These tend to produce each image as a rectangular framed canvas with its own colored background, rather than an isolated subject. Prefer 'illustration' or a specific technique: 'watercolor illustration', 'pen-and-ink sketch', 'ink wash', 'relief-etching', 'pastel drawing', 'woodblock print'.

• Avoid color-field or atmospheric phrasings in the style ('luminous backgrounds of violet, rose, and gold', 'set against jewel-tone fields', 'dreamlike rainbow atmosphere'). These instruct the image model to fill each image with colored atmosphere, producing framed compositions rather than isolated subjects. Describe only the linework, palette, and technique of the subjects themselves.

• Do not describe an aged, weathered, cracked, or textured surface, ground, wall, panel, or paper that the whole artwork sits on ('on aged wood', 'cracked fresco wall', 'aged parchment surface'); name the art tradition(s) or style(s) instead ('fresco-style illustration'). Texture that belongs to a subject's own material is fine ('a weathered bronze shield', 'a cracked ceramic vase').

• No captions, labels, or annotations. Text that is part of the depicted object is fine (e.g. 'STOP' on a stop sign, 'EXIT' on an exit sign).

• No grid lines, borders, frames, or separators.

• No overlapping or collage-style arrangements.

• Do not connect the subjects to each other or give them shared physical elements: no wires, cords, chains, ropes, ribbons, vines, or threads running between subjects, no frame or banner they share, no phrasing like 'connected by' or 'strung together', and no single continuous line or tube forming multiple subjects. Each subject must be drawable in complete isolation; connections inside one subject (a chain on an amulet, laces on a boot) are fine.

• No dramatic/long drop shadows (subtle shadows are fine).

• Image descriptions should describe WHAT to depict, not where to position it.

• Each image is ONE isolated subject, not a scene. Describe the subject with its pose or action and anything it directly holds, rides, or interacts with, but not the surrounding setting, environment, landscape, or sky. For a single composed scene (a figure set within an environment), use generate_illustration instead.

• Do not use size words (large, tiny, small, etc.) on the overall image subject (e.g. 'a large elephant', 'a tiny mouse') - all images are produced at the same size. Size words on details within the image are fine (e.g. 'a plate with a small insignia').

• Maximum 18 images per generation. Do not put the image count in the style.

• Subjects must be distinct: entries that differ only in case, punctuation, or spacing count as the same subject and the call is rejected. Explicit filenames must be distinct too (a different extension alone is not distinct).

• The style must actually describe a visual style, and each subject must name a drawable subject; text that does not is rejected.

• Style description max length: 500 characters. Image description max length: 200 characters each.

• Size: width and height are separate parameters, each 256 to 512 pixels when given. Both given is an exact box; one given fixes that axis and the other hugs each subject (so images in the set differ on it, and it may fall below 256); neither given delivers native resolution, which is also the path to larger images.

• Sizing: relative (the default) keeps the sizes the model gave the subjects in relation to one another, one scale for the whole set; fit scales each subject on its own to fill its frame less the margin, the icon-set convention, giving up relative size and enlarging subjects smaller than the frame (the result says by how much). Both can be changed later with edit_image_set.

• Icons: the subject count sets how large a batch's icons can later be exported with export_icons, crisp at every density: about a 136px base size with 13 to 18 subjects, 160 with 10 to 12, 180 with 7 to 9, 192 with 5 or 6, 256 with 4 or fewer. Every generation result states its batch's own crisp base size.

• If the style check returns a suggested cleanup, show the user the specific changes and get their confirmation, then resubmit the approved prompt with validation set to "skip" so it generates exactly as approved (resubmitting without "skip" re-runs the check and may return further suggestions). See the validation parameter for when to use "skip" and "auto-apply".

• If a "Rate limit exceeded" error is returned, wait the suggested number of seconds before retrying. Do not retry immediately.