openrouter-comfy-node

ComfyUI custom node for OpenRouter chat completions with:

• OpenRouter-only routing (Groq removed)
• Vision input from real ComfyUI IMAGE tensors (single image or batch)
• Reasoning effort selection (none, low, minimal, medium, high, xhigh)
• Robust JSON extraction from responses

Node key:

• ArrakisOpenRouterNode
• Display name: Arrakis OpenRouter (Vision + Reasoning)

Inputs

Required:

• api_key: OpenRouter API key
  • You can pass the literal token or an environment variable reference, e.g. $OPENROUTER_TOKEN or ${OPENROUTER_TOKEN}
• system_prompt: Instruction prompt (default enforces JSON-only output)
• user_prompt: Main user prompt
• user_image: Optional IMAGE input (connect image node directly)
  • Single image: sends one image_url message
  • Batch image: sends one image_url message per image in the batch
  • Images are encoded internally as data:image/png;base64,...
• reasoning_level: none, low (default), minimal, medium, high, or xhigh
• max_tokens: default 0 (unlimited / omit from payload)
• model: Default x-ai/grok-4.1-fast

Optional:

• custom_parameters: Raw JSON merged into request body (except reserved fields)
• timeout: Request timeout in seconds
• max_retries: Retry attempts for retryable errors
• enforce_json_output: Injects a strict JSON-only suffix into the system message (default: true)

Message Order

The node builds request messages in this order:

1. system message
2. user message(s) from user_prompt
3. image parts from user_image are appended to the same final user message content array (after text)
4. extra messages from custom_parameters.messages (if provided)

This keeps prompt text before image context and matches OpenRouter multimodal guidance.

Reasoning

By default, requests include:

"reasoning": {
  "effort": "low"
}

You can override via reasoning_level or fully override using custom_parameters.reasoning. If reasoning_level is none, the node omits the reasoning field from the payload.

Example custom_parameters

{
  "response_format": {
    "type": "json_object"
  }
}

Notes

• max_tokens from custom_parameters is ignored; use the dedicated max_tokens input.
• No dedicated inputs for temperature, top_p, top_k, etc. If needed, pass them via custom_parameters.
• For strict structure, prefer response_format with JSON schema in custom_parameters.
• Non-supported parameters may be ignored by model/provider routing.
• For IMAGE batches, the node samples at most 3 frames (first, middle, last).
• Images over 1 MP are resized before encoding to reduce payload size and cost.
• This repository contains a ComfyUI Python custom node. It is not a standalone Node.js app. Running python __init__.py now prints a local smoke-test status to help diagnostics.
• The node now includes a seed input. Keep the same value to reuse cache when inputs are unchanged, or change seed to force a fresh API call.