Hosted LLM API (Secure)

Local GGUF models are great until the task needs a bigger brain. PromptGenerateAPI is the pack's escape hatch: instead of loading a model into your VRAM, it calls a hosted LLM - ChatGPT or DeepSeek - to write your prompts. No local weights, no llama-cpp-python, no VRAM cost. The trade is an API key and a network call, but for hard prompt-writing jobs the frontier models earn their keep.

What it is

It's the API-backed member of the pack's prompt-generation family. The README points at DeepSeek specifically because platform.deepseek.com hands out 10M free tokens, which is a genuinely great way to try this node for nothing. It doubles as a simple chat node - there's an explicit toggle for that - so you're not locked into prompt generation if you just want to ask questions.

The inputs that matter

• model_name - dropdown. ChatGPT-3.5, ChatGPT-4, DeepSeek, plus the raw model IDs (gpt-3.5-turbo, gpt-4-0613, gpt-4-1106-preview, glm-4, and friends). Pick the provider, then the model.
• api_key - your key for the chosen provider. This is a multiline STRING; paste it in. Keep it to the model you actually selected - a DeepSeek key won't work on a ChatGPT model and vice versa.
• chat_type - BOOLEAN, default true. True = Prompt Generator mode; false = Simple Chat. The toggle that decides whether you get a crafted prompt or a conversational answer.
• description - context for the prompt generator. Describe the image, the style, the vibe you're after; this is what the API turns into a prompt.
• question - your actual request. In chat mode this is just the message; in prompt mode it's the instruction.
• context_size - how much conversation context to keep, default 5. The README frames this as a memory knob; more context means the model remembers more of the exchange.
• seed - for reproducible outputs where the API supports it.

Output is a single STRING.

How it works

Under the hood it's an OpenAI-compatible API call - the same wire format regardless of whether you're talking to OpenAI or DeepSeek. Your description and question get packaged up, sent off, and the returned text comes back as a STRING you can wire straight into a CLIP Text Encode or a text display.

Install

Just the pack, nothing else:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI_VLM_nodes

Restart ComfyUI (or Manager → "VLM Nodes"). No llama-cpp-python needed for this node - that's the whole appeal - and Python 3.9+ still applies.

Gotchas

The failure modes here are API-shaped, not GPU-shaped. Wrong or missing key → auth error; check the provider matches the model_name. No network → timeout; it's a hosted call, so ComfyUI needs internet. And a real heads-up: this node sends your description and question to a third party. If you're captioning or referencing images with sensitive content, the local models are the privacy-respecting option - that's a legit reason to reach for LLMSampler instead.