Conversational

Conversational is the closest thing this pack has to a chatbot. You send a message, it replies, and - the part that makes it a conversation rather than a one-shot - it keeps a history so it remembers what you said earlier. Inside a ComfyUI graph that's genuinely handy for building interactive workflows: a node that collects user input, responds, and hands the updated thread back so the next run picks up where the last one left off.

The node runs a Hugging Face text-generation pipeline under the hood, defaulting to microsoft/DialoGPT-medium. DialoGPT is Microsoft's conversational variant of GPT-2, trained specifically on Reddit dialog data - it's a real chat model, small enough (~345M parameters) to run on modest hardware. The node formats your message plus the stored history as a chat, generates a response (capped at 128 new tokens), and returns both the reply and the updated history.

The inputs that matter

• user_message - the current thing you want to say, a multiline string.
• history - a JSON array of previous turns, default []. This is how memory works: each run, the node reads it, appends your new message, and returns an updated version. Wire the updated_history output back into this input to keep the conversation going.
• model_name - Hub model ID as a string, default microsoft/DialoGPT-medium.

Two outputs: response (a STRING - the model's reply) and updated_history (a STRING JSON list of the whole thread so far).

How to install it

Standard pack install - ComfyUI Manager (search "ComfyUI-Transformers") or:

cd ComfyUI/custom_nodes
git clone https://github.com/kadirnar/ComfyUI-Transformers

Restart ComfyUI. DialoGPT-medium is about 1.4GB on first download - one of the chunkier first loads in the pack - and requirements.txt covers the rest of the stack.

Where people get burned

Manage your expectations about the model, because this is 2019-era GPT-2 tech. DialoGPT will hold a short back-and-forth and occasionally be funny, but it's nowhere near modern LLM quality - it rambles, forgets context, and says odd things. Also note the output shape: the pipeline returns the full chat including your input, and the node extracts the last assistant turn - if history ever comes back malformed JSON, the node silently starts a fresh conversation. It's a fun, self-contained chat loop for a workflow, and a nice way to add an interactive text element without leaving ComfyUI. Just don't pitch it against anything released in the last few years.