Gemini 3.1 Flash-Lite

The VLM (vision-language model) nodes in this pack are where the workflow ideas get interesting. You generate an image, wire it into this node alongside a question, and get a text answer back. The practical uses are the ones that turn ComfyUI into a self-correcting loop: auto-caption your generations for training data, check every output against a spec ("does this character have exactly six fingers?"), or route images - send them to a different branch based on what the model reports.

This node is Gemini 3.1 Flash-Lite looking at pictures, hosted by Google and paid through your BizyAir key. It's the cheap end of the vision lineup, which is exactly why it's the one you'd wire into a loop that runs on every batch. Same zero-local-footprint story as all the Gemini nodes: no weights, no VRAM, just API calls.

Inputs and outputs

• images - required, and it's an auto-growing input (COMFY_AUTOGROW_V3), so you can hand it several images in one call. The hosted model accepts up to 900 per request - you won't hit that, but knowing it grows saves you from wiring dozens of Load Image nodes.
• system_prompt - required. The framing: "You are a strict quality reviewer."
• user_prompt - required. What you want it to say about the images.
• detail - combo, default low. Set to high when you need it to actually read small text or count objects; low is faster and cheaper and fine for gist-level checks.
• temperature - 0 to 2, default 1.
• max_tokens - default 32768.
• enable_thinking - default false. Turn it on for reasoning-heavy checks.

Outputs: result (STRING) and request_id.

How it works

Serialize images plus prompts to the gemini-3.1-flash-lite-official/vision endpoint, poll, get text back. Images upload as part of the request, so bigger inputs mean slower and pricier calls - another reason detail defaults to low.

Gotchas

• "It said nothing" - the result is a string; you need a text preview node to see it, and if your graph has no display, it genuinely produces nothing visible.
• Detail matters for detail. For finger-counting or text-checking, low will hallucinate confidence. Set detail to high for that specific node.
• Failure with skip_error on returns the error text as the result - read the console for the real message.

Install

With the pack:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart ComfyUI, turn BizyAirPlus ON with a valid key, and wire your last Save Image into it. A cheap vision node that checks your own output is one of the best workflow upgrades you can make - it's the difference between shipping a batch with six fingers and catching it in the graph.