How it works
Two-phase tool. First the agent passes a reference image (URL or base64) plus what to vary; the tool runs a vision pass, produces N concept prompts, and returns them immediately. The agent then renders each concept via generate_image (one call per variant, supports multi-aspect). Perfect for taking a winning ad and producing 10 variants for A/B testing without losing the visual DNA.
Inputs
Returns
{ concepts: [{ name, imagePrompt, cta, headline, body }], requestedAspectRatios } — agent then calls generate_image per concept.Examples
5 variants of a winning ad, square + Google Ad sizes
Plug Motors had a winning Tesla creative. Agent generates 5 variants for next week's test, each in both sizes.
{
"referenceImageUrl": "https://cdn.claivra.com/file/claivra/.../tesla-winner.jpeg",
"count": 5,
"varyOptions": [
"vehicle",
"background",
"headline"
],
"aspectRatios": [
"1:1",
"1.91:1"
],
"brandId": "aBmbZETe9gUZhwXHCE02",
"notes": "Keep the green spotlight + reflective floor — those are the brand cues that work."
}Related tools
generate_imageGenerate an ad creative image
Generate a single (or multi-aspect-ratio batch) ad creative from a director-style prompt, with optional brand kit + logo composite.
edit_imageEdit an existing image with an instruction
Apply a natural-language instruction to an existing image — swap headlines, change colors, remove text, replace backgrounds — without regenerating the whole composition.