🤖 Gemini API

Somewhere in a Qwen workflow you usually want a plain-language description of the input image - and the model that's best at that is often not the one generating the image. ArchAi3D Gemini is a vision-language API node: it sends up to four images and a prompt to Google's Gemini, and returns the answer as a string you can wire anywhere a prompt string goes. It's the kind of node that slots in before your Qwen encoder to auto-generate the scene description, or after a generation to caption what you made.

The inputs are straightforward. prompt (default "Describe this image in detail.") is what you ask; model defaults to gemini-2.5-flash, which is the sensible default - cheap, fast, and plenty smart for describing interiors. The dropdown includes gemini-2.5-pro, gemini-2.5-flash-lite, the 2.0 line, and a few 3.x previews. system_prompt lets you set a persona ("You are an expert interior designer...") to steer the response. image1 through image4 are plain IMAGE inputs, so you can feed a batch of reference shots.

The knobs that matter

• api_key - your Google AI Studio key. The node saves it to its own config file after the first successful call, so you only paste it once.
• temperature, top_p, top_k, max_tokens - standard sampling controls. Defaults (1.0 / 0.95 / 40 / 8192) are fine for captioning.
• thinking_budget (0-24576, default 0) - this is the interesting one. Turn it on for the 2.5/3.0 models and they get a token budget to reason before answering. For a "describe this room precisely" job it noticeably improves accuracy; for simple captioning it just costs latency.
• use_cache (default on) - the node hashes its inputs and reuses the previous response if nothing changed. In a looping workflow this saves you paying for the same API call every iteration. Leave it on.
• model_override - connect the pack's ArchAi3D Gemini Model node here to switch models from another part of the graph, instead of hard-coding the dropdown.

Outputs and install

One output: response, a STRING. Wire it to a ShowText/display node to read it, or straight into a prompt-builder input.

Installation is the pack install, but with a real prerequisite: this node needs the google-genai SDK, and if it's missing the node doesn't even register - the pack silently skips it (the import is wrapped in a try/except). So:

cd ComfyUI/custom_nodes/
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen
pip install -r requirements.txt   # includes google-genai>=1.0.0

or install pip install google-genai into your ComfyUI environment directly, then restart. ComfyUI Manager's "search ArchAi3d Qwen" route handles it all too.

Where people get burned

Missing node = missing SDK, most of the time. If the node never appears after install, check that google-genai actually landed in the right Python environment. Second: this is a paid API - every run costs money, and use_cache off means every run costs money. There's a free tier, but a batch workflow re-querying on each iteration will surprise you on the bill. And remember thinking_budget isn't supported by every model in the list; models that don't do reasoning ignore it or error, so keep it at 0 unless you're on a 2.5/3.x model and actually need the reasoning.