Gemini 3 Flash

An LLM inside ComfyUI sounds like a gimmick until you've built a workflow that needs one. The honest use cases are real: prompt expansion - feed it "neon cyberpunk street, rain" and get a full paragraph you hand to your image model; prompt translation; or a loop that rewrites a prompt and re-queues until the output is right. This node is the official Gemini 3 Flash running on Google's side, routed through BizyAir, which means no local model, no VRAM, and no quantized-LLM rabbit hole. It costs cloud credits and needs network - that's the whole deal.

The display name says "Official," and it means it: this isn't SiliconFlow's own model dressed up as Gemini, it's the real thing behind an API call. The key that pays for it is your BizyAir key, so one account covers your image model and your text model.

Inputs and outputs

• system_prompt - required. The role and rules. "You are a prompt engineer. Expand user ideas into detailed image-generation prompts."
• user_prompt - required. What you actually want done.
• temperature - 0 to 2, default 1. Higher is more random; for structured prompt expansion, people usually go lower, not higher.
• max_tokens - default 32768. Plenty of room for long rewrites.

Optional extras worth knowing: enable_thinking (default true on this model - let it reason before answering, it's usually worth it for rewriting tasks), enable_search (default true - lets the model pull live info, which matters if your prompts reference current events). Then the pack-wide api_config override and skip_error.

Outputs: result (the model's text) and request_id. Wire result into a text preview or a string node - in a default graph you'll want a Show Text-style node to actually see it.

How it works

Standard BizyAir flow: serialize system prompt, user prompt, and sampling settings, submit to the gemini-3-flash-official/large-language-models endpoint, poll, and return the text. The api_config input lets a settings node override base URL and key per request; skip_error makes a failed call return a placeholder instead of killing the workflow.

Gotchas

• Thinking + search are on by default. That's slower and pricier than a bare completion. If you just want a fast rewrite, flip enable_thinking off.
• Text output disappears. The result is a STRING, not a pretty preview. If your workflow "produces nothing," check you actually have a node showing the string.
• Red placeholder / error string on failure with skip_error enabled - 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, set your key. Then wire result into your next text encoder and let Gemini 3 Flash write your prompts - that's the workflow this node was built for.