OpenRouter LLM

OpenRouter LLM is the node you reach for when you want a genuinely smart LLM - think Llama 4 Maverick, Qwen2.5-VL 72B, Kimi K2 - writing prompts or describing images in your workflow, but you don't want to give up 8GB of VRAM (or any of it) to a local model. It calls the OpenRouter API from inside the graph, so the heavy lifting happens on their servers. The catch is the obvious one: it needs an API key and an internet connection, and the free-tier models are rate-limited. If that trade sounds fine, this is the smoothest way to get there.

How it works

The node posts your prompt (and any attached images) to OpenRouter's https://openrouter.ai/api/v1/chat/completions endpoint with a bearer token. That's the whole mechanism - no local model, no llama.cpp, no VRAM. Images you connect get base64-encoded into the request, which is why this handles vision models like the Qwen2.5-VL entries in the dropdown without you doing anything special. For the handful of image-generation models in there (the Gemini image-preview ones), the node sets an image output mode and returns the picture as image_output; for normal text models it just returns text_output. The node even remembers your key in a local config.json, so you paste it once and leave the field blank afterwards.

The inputs that matter

• api_key - from openrouter.ai/keys. Free account is fine; most models in the dropdown are :free variants.
• text_model - dropdown of ~10 mostly-free OpenRouter models, plus an "Add Custom Model..." option where you type into custom_model. Default is google/gemma-3-27b-it:free, which is a solid prompt-optimizer.
• user_prompt - your actual request. Multiline.
• prompt_template - pick a .md template from the pack's Prompt folder (photography, image-to-prompt, person-feature extraction, video-frame transitions…) or leave "Custom" and write your own system_prompt.
• image_input_1/2/3 - optional images for vision models.
• enable_resize with target_width/target_height/resize_method - pads or scales images to a fixed size before upload. Only active when enable_resize is on; the author's own tooltips say so. Leave it off unless a model rejects odd-sized images.

seed exists, but read its tooltip before trusting it: some models honor it, and image-generation models like Gemini simply don't support a seed parameter. Set it for reproducible text, don't expect it to work everywhere.

Installing it

This pack ships with a bunch of unrelated nodes, so you get OpenRouter LLM as a bonus:

cd ComfyUI/custom_nodes
git clone https://github.com/dseditor/ComfyUI-ListHelper
# restart ComfyUI

Or use ComfyUI Manager: Install Custom Nodes → search "ComfyUI-ListHelper". Base dependencies (regex, accelerate, reportlab) install with the pack; the node uses requests, which ComfyUI already ships.

Where people get burned

The 429. Free OpenRouter models rate-limit hard, and the node surfaces it as a ❌ message in the output. If you're batch-processing hundreds of frames through a :free model, you will hit it - either pay for a cheap model or throttle your loop. Also remember this is cloud inference: your images leave your machine. If you're working on private client work, that matters more than the model choice.