LLM Message

Chat models don't take a single string - they take a list of messages, each tagged with a role: system (the instructions), user (what you're asking), assistant (what the model said before). This node is how you build that list inside ComfyUI. One node makes one message; chain a few together and you've assembled a whole conversation to hand to LLM Chat. It's also where you attach images, which is what makes vision prompts possible.

If you only ever want to fire a single prompt with no system message and no images, LLM Completion skips all this. Reach for LLM Message when you want a system prompt, a multi-turn setup, or images in the conversation.

How it works

Each node holds one message: a role and its text. To build a conversation, you chain - set the first node to system with your instructions, feed its messages output into a second node's optional messages input, set that one to user with the actual prompt, and so on. Each node appends its message to the growing list. The final messages output is the complete conversation, ready for the Chat node.

The image inputs are the other half of the trick. Because models like Gemini are multimodal, you can attach up to four images to a message and ask the model about them - describe this, caption that, compare these. That's how you wire a generated or loaded image into an LLM step.

The inputs and outputs

• role - system, user, or assistant. System sets behavior, user is your input, assistant represents prior model turns (useful for few-shot examples).
• text (multiline STRING) - the content of this message.
• messages (optional, LLM_MESSAGE) - the incoming list to append to. Leave empty on the first node in the chain.
• image, image_2, image_3, image_4 (optional, IMAGE) - up to four images to attach to this message, for vision-capable models.
• messages (output, LLM_MESSAGE) - the accumulated conversation.

Installing it

ComfyUI Manager → search comfyui-art-venture → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

then restart ComfyUI. Pre-installed on comfy.icu.

Common issues

Missing-node red means the pack isn't installed - Manager → "Install Missing Custom Nodes."

The mistakes here are about message structure, not the node itself. Attach images to a message but point the Chat node at a text-only model and the images are ignored - vision needs a multimodal model (Gemini, GPT-4o-class) behind the API node. Forget to chain the messages output forward and each node is an island; the Chat node only sees the message list it's actually wired to. And order matters: most chat APIs expect the system message first, then alternating user/assistant turns - build the chain in that order rather than trusting it to sort itself. Get the roles and the wiring right and this becomes the flexible front-end to every LLM call in your graph.