Modelverse Chat

This is the odd one in the pack and probably the most useful sleeper. Every other node here is a fixed model with a fixed job; Modelverse Chat is an OpenAI-compatible chat client that can talk to any model the Modelverse API serves. The model dropdown is dynamically fetched from the API - the brief shows bge embeddings, Doubao prose models, MiniMax, IndexTTS-2 text-to-speech, and "…+210 more" - so it's less a node and more a front door to the whole catalog.

What it is

A chat-completion node in the shape of a ComfyUI node. You give it a user prompt, optionally a system prompt, optionally an image or files, and it returns the model's response as a response STRING. In practice people use it three ways: as a prompt-engineer in the loop (ask the model to rewrite your prompt, feed the result into a generation node), as a free-form text tool alongside generation, or as a routing point where one node reaches many models.

The inputs that matter

• client - required, from the Modelverse Client / Secret Client node.
• model - the dropdown, populated from the API on first load with a curated fallback list if the fetch fails. Defaults to zai-org/glm-5.
• user_prompt - the question/instruction.
• system_prompt - optional persona or constraints ("You are a helpful assistant" by default).
• temperature (0.7), top_p (1), max_tokens (6000) - the usual sampling knobs. Lower temperature for more deterministic outputs.
• presence_penalty / frequency_penalty - push the model toward novelty or away from repetition.
• image_in - an IMAGE tensor for vision models. This is how you get a vision LLM in your graph.
• files - an OPENAI_INPUT_FILES input, fed from the OpenAI Input Files node. Note it says "files," plural - document context for models that accept it.
• response_format - text only, currently.

The single output is response (STRING). Whatever the model says, in one string, ready to wire into a prompt field elsewhere.

Installing and keying up

Standard pack install. ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ucloud/comfyui-plugin

restart ComfyUI. Deps are light (requests, pillow, numpy, openai, GitPython); the openai dependency is what powers the chat client. Get a ModelVerse API key from the UCloud console and enter it via the Modelverse Secret Client node (name in workflow, key in secrets.json) or a config.ini MODELVERSE_API_KEY=... entry.

Common issues

• Model list is short or stale - the list is cached after first fetch. If the catalog changed, restart ComfyUI or clear the cache so it refetches; if your key isn't valid the node silently falls back to the built-in list.
• Model doesn't support images - you wired image_in to a text-only model and got a refusal or a confusing reply. Pick a vision-capable model.
• "API key is not set" - empty client node. The eternal first check.
• Long responses truncate - max_tokens defaults to 6000 but some models have shorter context. Lower it or pick a bigger model.
• Billing by token - chat is cheaper per call than video, but the meter still runs. Batch your prompt work.

The genuinely clever workflow: chain this node's response string into a Flux or video node's prompt field, and you've built a "director LLM" that writes the prompts your generators execute. That's the real reason this node exists, and it's worth having just for that.