Google Gemini 3.6 Flash (Official Edition)

Read the images tooltip on this node and stop for a second: up to 3,000 images per request, PNG/JPEG/WebP/HEIC/HEIF, 7 MB per inline image. That's not a captioning node - that's a batch-vision node. Hosted Gemini 3.6 Flash, billed through your BizyAir key, and it's the newest vision entry in the pack. Combined with response_mime_type and response_schema for structured output, this is the node you'd build a whole auto-captioning or batch-quality-report pipeline around.

The use case that makes this interesting is the one the older vision nodes can't do well: analyzing a collection. Caption an entire dataset folder in one call. Look at ten reference frames and ask which ones violate the style guide. Get back a JSON array of per-image verdicts instead of a wall of prose you then have to parse.

Inputs and outputs

• images - required, up to 3,000 (7 MB per inline image, 30 MB per Cloud Storage image). Note this one is typed IMAGE, not autogrow - but the hosted limit is what matters, and it's enormous.
• prompt - required. Up to 1,048,576 input tokens.
• system_instruction - optional.
• stop_sequences - optional, up to 5.
• response_mime_type - combo, default text/plain. Set to application/json for structured answers.
• response_schema - optional JSON schema, only effective when the mime type is application/json.
• seed - default 0 (-1 = random).
• max_output_tokens - default 8192, up to 65,536.

Outputs: result (STRING) and request_id.

How it works

Images and text serialize up to the gemini-3.6-flash-official/vision endpoint, the cloud analyzes, and the text returns via standard polling. The key comes from BizyAirPlus settings or BIZYAIR_API_KEY; api_config overrides per request. With skip_error on, a failure returns the error text as the result rather than killing the workflow.

The honest take

Two practical notes. First, the structured-output pairing matters even more here than on the LLM version: asking for a JSON verdict per image is the difference between "the model said the batch is mostly fine" and "image 4 fails the spec, image 7 passes" - and the schema is the thing that enforces the shape. Second, mind the 7 MB per-image cap: for large source files, downscale or re-encode before wiring them in, or you'll get rejected before the model ever sees them. The 3,000-image ceiling is generous enough that the practical limit is almost always your upload time and your credit balance, not the model.

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, confirm BizyAirPlus is ON and keyed, and point it at a folder of images with a JSON schema in hand. It's the closest thing this pack has to a batch quality inspector, and it's worth the credits for that job alone.