✨ NVIDIA Prompt Generator (Text) ✨

Let's get one thing straight before anything else: this node doesn't run a model. It's a thin OpenAI-SDK client pointed at NVIDIA's hosted NIM API. Your text goes up to the cloud, Llama 3.3 70B (or a 405B, or Gemma) writes you back a prompt, and your GPU never wakes up. If you want a genuinely good LLM to turn "majestic lion at sunset" into the kind of structured, camera-and-lighting instruction a 2026 model actually responds to - but you'd rather not babysit a GGUF in Ollama or give up VRAM mid-workflow - this is the API path to exactly that.

It's the honest version of that trade, too. The community default for in-graph prompt enhancement is a small local model, because local is uncensored, offline, and free per call. Going through NVIDIA's cloud buys you frontier-chat quality at the price of the four things every API node costs: per-call money, your prompts leaving the machine, NVIDIA's content filter, and the fact that a cloud model can be retired out from under you. Fine for SFW prompt drafting. Wrong tool if "uncensored" is the point.

How it works. The node builds a system + user message and calls the OpenAI-compatible chat completions endpoint at integrate.api.nvidia.com/v1. On top of that the author added the plumbing that makes cloud nodes usable: a one-second minimum between requests, retries with a configurable delay, and an in-memory cache keyed on model, prompt, and your sampling params. The genuinely thoughtful bit is model-quirk handling - Gemma and the Phi models don't support system roles or penalty parameters, so for those it merges the system prompt into the user turn and silently drops the penalties instead of erroring.

The inputs that matter. You'll actually touch maybe five of these:

• api_key - masked password field, starts empty. This is what bites people first.
• model - 12 choices, default meta/llama-3.3-70b-instruct. The -instruct/-it suffixes are the instruction-following variants.
• prompt - multiline, with dynamicPrompts enabled, so wildcards work.
• system_prompt - the default persona ("expert AI art prompt engineer") is genuinely decent; edit it if you want a different output shape.
• max_tokens (16–8192), temperature, top_p, use_cache, and an optional seed (-1 = unseeded).

The penalty sliders and retry knobs are there if you want them. You probably won't touch them.

The output is one text STRING. Wire it into any text input - CLIP Text Encode's positive prompt is the classic move, but a Save Text node or another LLM downstream is just as happy.

Install is the standard two-lane story. ComfyUI Manager: search "ComfyUI-NvidiaVision" and hit install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/theshubzworld/ComfyUI-NvidiaVision
cd ComfyUI-NvidiaVision
python_embeded\python.exe -m pip install -r requirements.txt   # only openai is declared

Restart ComfyUI and paste your NVIDIA API key from the developer portal into the node.

Where people get burned. Forgetting the key (the node's error message is basically "NVIDIA API key is required."). Expecting a retired model to still answer - the dropdown is a snapshot of NVIDIA's catalog, so if a call 400s, pick a live sibling model on build.nvidia.com. And assuming the cache persists: it's per-session memory, gone on restart. One more: the key lives in the workflow file, so don't share a saved workflow with your key in it. It's an API-wrapper node from exactly the category that got weaponized once (LLMVISION); this pack is small, open, and only phones NVIDIA, but reading what you install is the habit that keeps you safe.