Imagegen

Image Generation & Editing

Generate or edit images through ClawRouter. Payment is automatic via x402.

Shortcuts:

• Slash: /cr-imagegen <prompt> [--model=<alias>] [--size=1024x1024] [--n=1] (/imagegen still accepted in chat for backward compatibility)
• Partner tool: blockrun_image_generation (LLM-callable) / blockrun_image_edit (inpainting)

Generate an Image

POST to http://localhost:8402/v1/images/generations:

jsonCopy code
{
  "model": "google/nano-banana",
  "prompt": "a golden retriever surfing on a wave",
  "size": "1024x1024",
  "n": 1
}

Response:

jsonCopy code
{
  "created": 1741460000,
  "data": [{ "url": "http://localhost:8402/images/abc123.png" }]
}

Display inline: ![generated image](http://localhost:8402/images/abc123.png)

Model Selection

Choosing a model:

• Default → nano-banana
• "high res" / "large" → banana-pro
• "photorealistic" / complex scenes → gpt-image-2
• "flagship quality" / reference image → seedream
• "budget" / "cheap" → cogview
• "editable" / "inpainting" → gpt-image (only edit-capable model)
• "artistic" / flexible content → grok-imagine
• "grok style" → grok-imagine or grok-imagine-pro

Choosing a size:

• Default: 1024x1024 (the only size every model accepts)
• Portrait: 1024x1536 (gpt-image / gpt-image-2) or 1728x2304 (seedream)
• Landscape: 1536x1024 (gpt-image / gpt-image-2), 1344x768 (cogview), or 2048x1024 / 1280x720 (seedream)
• High-res: 2048x2048 / 4096x4096 with banana-pro; 2848x1600 with seedream
• The gateway validates size per model BEFORE payment and rejects unknown ones — do not invent sizes outside each model's list above

Edit an Existing Image

POST to http://localhost:8402/v1/images/image2image:

jsonCopy code
{
  "model": "openai/gpt-image-1",
  "prompt": "make the background a snowy mountain landscape",
  "image": "https://example.com/photo.jpg",
  "size": "1024x1024",
  "n": 1
}

ClawRouter automatically downloads URLs and reads local file paths — pass them directly, no manual base64 conversion needed.

Optional mask field: a second image (URL or path) that marks which areas to edit (white = edit, black = keep).

Response is identical to generation:

jsonCopy code
{
  "created": 1741460000,
  "data": [{ "url": "http://localhost:8402/images/xyz456.png", "revised_prompt": "..." }]
}

Supported models for editing: openai/gpt-image-1 only ($0.02)

Example Interactions

User: Draw me a cyberpunk city at night → POST to /v1/images/generations, model nano-banana, prompt as given.

User: Generate a high-res portrait of a samurai → POST to /v1/images/generations, model seedream, size 1728x2304.

User: Edit this photo to add a sunset background: https://example.com/portrait.jpg → POST to /v1/images/image2image, model gpt-image, image = the URL, prompt = "add a warm sunset background".

User: Change the background in my image to a beach (attaches local file) → POST to /v1/images/image2image, image = the local file path, prompt describes the change.

Notes

• Payment is automatic, and which rail it uses depends on how ClawRouter was started: an x402 USDC micropayment from the user's wallet, or a draw on account credit if a BlockRun API key is configured. Either way the agent does nothing.
• If the call fails with a payment error, check GET http://localhost:8402/health and read authMode before telling the user how to fix it: wallet → fund the wallet at blockrun.ai; api-key → top up account credit at user.blockrun.ai/dashboard/credits. Naming the wrong one sends the user to a page that cannot fix their error.
• Google models may return base64 internally — ClawRouter uploads automatically and returns a hosted URL
• OpenAI image models enforce OpenAI content policy; use nano-banana or grok-imagine for more flexibility
• Image editing is only available with gpt-image-1; generation supports all listed models