// Generate images for the user's vault — pick the right provider, gather reference images from notes when relevant, write the result to assets/generated/, and surface the markdown reference. Use this when the user asks for any picture, illustration, mood image, poster, or visual companion to a note.
| name | image-gen |
| description | Generate images for the user's vault — pick the right provider, gather reference images from notes when relevant, write the result to assets/generated/, and surface the markdown reference. Use this when the user asks for any picture, illustration, mood image, poster, or visual companion to a note. |
You have one image-generation tool: generate_image. It routes to one of three providers (gpt-image-2, Gemini Nano Banana, Seedream 4.5) and writes a PNG to assets/generated/<YYYY-MM>/<id>.png inside the vault, with a sidecar JSON capturing the prompt and references for reproducibility.
The tool is registered by Lumina's plugin. The provider's API key must already be configured in AI Settings → Image Models — surface that path in your error message if you hit a "no API key" error.
Trigger on intent, not specific words. Plausible signals:
Don't invoke when:
Pick the provider parameter based on what the request is for and what is configured. When Lumina tells you a configured provider in the user message wrapper, use only that provider unless Lumina explicitly lists another configured provider. Never switch to an unconfigured provider.
| Use this provider | When the request is… |
|---|---|
openai-image (gpt-image-2) | Photorealism is essential, the user explicitly asks for OpenAI / DALL-E / GPT, or Lumina says OpenAI is the configured image provider. |
google-image (Nano Banana) | General-purpose, iterating with reference images, anything where consistency across multiple frames matters, when Google image generation is configured. |
bytedance-image (Seedream 4.5) | The image needs to render readable text, dense typography, or poster-style layouts, and Seedream is configured. The language the user typed in is not by itself a visible-text requirement. |
When the user doesn't specify, don't invent a different provider. Use the provider Lumina identified as configured in the wrapped prompt; otherwise omit provider and let Lumina route to the first configured image provider.
Preserve the user's explicit visual constraints. Medium, region, era, genre, culture, composition, subject, mood, palette, and style descriptors are part of the image request — do not replace them with a nearby default style unless the user asked for that. Treat readable text separately: if the user asks for visible text, preserve the requested text and language; otherwise avoid readable text, letters, captions, labels, and speech bubbles.
Reference images are file paths. generate_image reads those files and sends their bytes to the image-generation provider; the chat model does not need vision capability for the references to affect generation.
Most requests do not need automatic reference search. Prefer explicit or high-confidence references:
@ or named directly.Only search for candidates when the user asks for an existing style/image without naming a specific file ("like my earlier cyberpunk cover", "match this project's visual style").
Workflow when reference search is actually needed:
glob to find image files in a narrow scope: current note folder first, then assets/generated/**, then user-named subdirectories.grep to find notes, embeds, filenames, and sidecar metadata mentioning the subject/style the user described.read on a small candidate set and extract visual traits that improve the prompt: subject, composition, palette, lighting, material, style, typography, and framing.read to judge what images depict. Choose only from explicit references or textual evidence: filename, path, nearby note text, embeds, recency, and sidecar prompt/metadata.reference_images.Cap at 3 references. Nano Banana stops paying attention beyond that, and gpt-image-2 / Seedream don't gain much either. If you have more candidates, pick the explicit/high-confidence ones first.
Don't fabricate paths or visual facts. If you can't find a reliable reference file, ask the user to point you at one or proceed without references.
Use aspect_ratio to match the use case:
16:9 — banners, hero images, desktop wallpapers, presentation slides9:16 — mobile/vertical stories, posters4:3 / 3:4 — when the user mentions photo / paper proportions1:1 — default when nothing else fits, icons, avatars, square thumbnailsIf the user doesn't say, infer from context — a "header image" is 16:9, an "icon" is 1:1, a "phone wallpaper" is 9:16.
The tool returns the vault-relative path, and Lumina renders the generated image in chat automatically. Do not repeat the image as markdown in the chat reply unless the user explicitly asks for the markdown reference. If useful, mention the saved vault-relative path in plain text.
If the user asked for the image to land in a specific note, edit the note to insert it (use the edit tool). Don't leave the image dangling — the whole point is integration with their vault.
If the user is iterating ("more X", "less Y"), pass the previous generation's absolute path back in reference_images so the next call carries the visual continuity.
read on the saved file.