Image Generation

Switch the generate_image tool on or off and set its default model, limits and output folder. The image models themselves are configured per provider under Providers → Image models — this tab can enable, disable and rename them, but new models are only added there.

URL: /settings?tab=imageGeneration

All changes apply from the next message — no restart. Background tasks pick them up as well: new tasks get the updated tool, and a running task that still holds the old tool is checked against the current settings on every call.

Enabled

Master switch for image generation. Default: on, so existing setups keep working.

When off, the agent gets no generate_image tool (in chat and in background tasks), the system prompt has no Image generation models block and no generate_image line in <available_tools>, and the built-in image-generation skill disappears from <available_skills> (it declares requires_toolsets: [generate_image]). Image models stay configured under Providers, so switching back on restores everything.

Even when on, the tool only exists while at least one image model is enabled on an enabled provider.

{ "imageGeneration": { "enabled": true } }

The remaining fields are shown while image generation is on.

Default image model

Image model used by generate_image when the agent does not pick one. The dropdown lists the enabled image models of all enabled providers; text models never appear here, and image models never appear in text-model dropdowns. This is the only place where the default is set.

First enabled image model (empty value) picks the first enabled image model. If the selected model is removed later, the same fallback applies; disabling the model or its provider resets the setting.

{ "imageGeneration": { "defaultModel": "<providerId>:recraft/recraft-v4.1-vector" } }

Max variants per call

Upper limit for the tool's n parameter (number of images in one call). Default: 4. Range: 1 – 10. The limit is part of the tool definition the agent sees; a call with a larger n is refused before anything is generated. Each variant is a separate, separately billed request.

{ "imageGeneration": { "maxVariants": 4 } }

Axiom has no spending limit of its own. To cap image costs, set a limit on the provider's API key (e.g. an OpenRouter key limit or an OpenAI project budget). Images generated through a ChatGPT subscription cost nothing extra but count toward its Codex usage limits.

Output folder

Folder inside the workspace where images are saved. Default: images. Each image lands in a YYYY-MM-DD/ subfolder of it, together with its JSON sidecar. The value must be a relative path that stays inside the workspace — absolute paths and paths that leave the workspace via .. are rejected on save.

{ "imageGeneration": { "outputDir": "images" } }

Image models

Every image model configured on an enabled provider, with provider, model id and the model's description. Per model:

• Switch (top right) — enables or disables the model for the agent. Disabled models stay configured but are hidden from generate_image and the default-image-model dropdown; same as Disable / Enable on the Providers page.
• Edit (pencil) — display name and description. The description is what the agent reads to pick a model, so describe what each model is good at.
• Provider name — links to the provider on the Providers page.

New image models are added, checked and removed only under Providers → Image models.

See also

• Image Generation — how the tool, the system prompt block and the skill work.
• Providers → Image models — enabling models, availability check, paid test image.
• settings.json → imageGeneration