LLM <-> Character

Every ComfyUI power user eventually wants an LLM in the graph - to write prompts, describe an image, or riff on a character - and the friction is always the same: API keys, cloud accounts, privacy. LF_CharacterImpersonator is the LF Nodes answer, and its name is a small lie in the best way: it doesn't call any commercial API and needs no key. It talks to a local endpoint - by default http://localhost:5001/v1/chat/completions - the standard address served by Koboldcpp, which the pack's README explicitly names as its intended backend.

You give it a character_bio (a biography that shapes tone and content) and a prompt, plus temperature and max_tokens to control the reply, and it streams back a response as if the character wrote it. The author's own framing calls it a tool for "creative writing, role-playing scenarios, or generating dynamic content based on character traits" - so, LLM-driven character dialogue inside a generation workflow.

The inputs and outputs that matter

• character_bio - the character's backstory, personality, speech patterns. The more you put here, the more the output stays in voice. This is the input that makes the node what it is.
• prompt - what you ask the character. Single or multi-line.
• temperature - 0.1 to ~1.9, default 0.7. Higher = more chaotic text.
• max_tokens - cap on reply length, default 500.
• seed - reproducibility of the generation.
• url - defaults to the local Koboldcpp endpoint. Change it if you run something else compatible with the OpenAI chat-completions shape.
• image (optional) - an image that "can influence the output based on visual cues." This is how you get vision: point the node at a portrait and have the character react to it. Note this needs a multimodal backend (the README's sample stack is Llama-3-Instruct-8B with the LLaVA-Llama-3-8B projector) - a text-only model will ignore or choke on it.

Outputs are answer (the reply, ready for a display or prompt node), plus request_json and response_json if you want to inspect or forward the raw payloads.

The honest setup cost

The node itself installs trivially, but the backend is the real dependency: you need Koboldcpp (or an OpenAI-compatible local server) running on port 5001, with a model loaded that suits the task. That's a one-time setup - download Koboldcpp, drop in a GGUF, start it - and from then on the node just works. If you don't already run a local LLM, budget an afternoon; if you do, this is a five-minute hookup.

Installing it

The node comes from the LF Nodes pack:

cd ComfyUI/custom_nodes
git clone https://github.com/lucafoscili/comfyui-lf

or ComfyUI Manager → LF Nodes → install → restart. No Python deps beyond the pack's own. The pack is the frozen legacy repo (migrated to lf-nodes in early 2025), so the LLM nodes here are as they were last shipped.

Where people get burned

• Nothing running on localhost:5001 → connection refused, and the error can read like the node is broken when it's actually just that Koboldcpp isn't up. Start the server first.
• A text-only model with an image attached will misbehave or error. Match the image input to a vision-capable stack.
• This is the LLM node with the most character, but if you want a conversation rather than a one-shot impersonation, the pack's LLMChat node is the chat UI. Don't confuse the two.

For local, private, no-key LLM text inside ComfyUI, this is the friendliest on-ramp in the pack.