Call LLM Service 🐑

Everything else in the MieNodes LLM family is a preset around this node. CallLLMService is the raw one: you hand it a connector and some text, it sends that to the LLM, and it hands you back the model's reply as a string. No preset system prompt, no model-specific formatting - just a chat call you fully control. It's the node the pack's own "Call LLM Service" example workflow is built on.

Part of ComfyUI-MieNodes (ComfyUI_MieNodes), MieMieeeee's utility pack, under LLM Service Config.

Why you'd reach for it

The specialized generators (Kontext, Flux2, Ideogram4) bake in a system prompt tuned for one target model. That's great when you want their opinion on how to write for that model. But sometimes you just want to ask an LLM a thing inside your graph: translate this caption to English, summarize these tags, turn this one-liner into three sentences, describe what's in this image. That's CallLLMService.

It's also multimodal when your provider is. There's an optional image input, so if your connector points at a vision-capable model (Gemini, a MiMo Omni variant, and so on) you can pass an image and ask the model about it. That turns this node into a lightweight "describe this image" or "caption this" step without pulling in a separate captioner.

The KB's prompt-engineering doc is worth reading alongside it: on modern LLM-encoded diffusion models the winning move is to write your prompt as a clear instruction, and this node is the most direct way to have an LLM produce that instruction - or to translate, expand, or restructure anything else on the wire.

The inputs and outputs that matter

Required:

• llm_service_connector - from any of the Set...LLMServiceConnector nodes (DeepSeek, MiMo, General, etc.).
• input_text - the prompt/message you're sending. Multiline, so write as much as you need.

Optional, the standard sampling knobs plus vision:

• temperature (default 0.7) and top_p (default 0.9) - creativity vs determinism.
• max_tokens (default 512) - cap on the reply length. Bump it if you're asking for long output; 512 is short.
• seed (default 0) - for repeatability where the provider honors it.
• image (IMAGE) and image_detail (auto/low/high) - pass an image to a vision model, and hint how carefully it should look.

Output is a single response (STRING) - the model's text. Wire it into a text-encode node, a SaveAnythingAsFile, or a show node to eyeball it.

Installing it

ComfyUI Manager → search ComfyUI-MieNodes → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes

then restart. No model download - this calls a remote LLM. You do need a working connector, so set one of those up first. Nodes appear under the 🐑 MieNodes menu.

Common issues

Truncated replies. max_tokens defaults to 512. If your output is getting cut mid-sentence, that's almost always the cause - raise it.

The image input does nothing. That's expected if your connector points at a text-only model. Vision needs a multimodal model (Gemini, a MiMo/MiniMax Omni variant). Passing an image to a text model either errors or is silently ignored, depending on the provider.

Errors before any generation. That's the connector, not this node - bad key, wrong endpoint, or unknown model id. Fix it on the Set...Connector node upstream.

It's a chat call, so it costs tokens. Every run bills against your provider. If you're iterating on a big graph, remember this node fires each queue unless the inputs are unchanged and the result is cached.