Gemini 3.5 Flash

The 3.5 Flash VLM is the vision sibling of the most controllable text node in the pack - and it inherits all the control. It looks at your images, answers about them, and unlike the 3.1 vision nodes it hands you seed, top_p, top_k, and stop_sequences. That makes it the vision node to use when your captioning or quality-checking loop has to be reproducible and well-behaved: same seed, same images, same question, same verdict.

The workflow idea is the same as every VLM node in this pack - generate, ask the model what it sees, act on the answer - but the extras make it better suited to loops that parse the result. stop_sequences is the quiet MVP there: have the model end its answer at a marker, and your downstream string-splitting never has to guess where the analysis stops.

Inputs and outputs

• images - required, auto-growing (COMFY_AUTOGROW_V3). One or more images for the model to analyze.
• prompt - required. The question about the images.
• system_instruction - optional. The framing and rubric.
• stop_sequences - optional. Halt generation at a marker.
• seed - default 0, for reproducible answers.
• temperature - 0 to 2, default 1.
• top_p - 0 to 1, default 0 (unset convention, like the LLM node).
• top_k - default 0, same.
• max_output_tokens - default 32768.

Outputs: result (STRING) and request_id.

How it works

Images and text serialize up to the gemini-3.5-flash-official/vision endpoint, the cloud runs the analysis, and the text comes back via the standard BizyAir polling loop. Your key comes from BizyAirPlus settings or BIZYAIR_API_KEY unless overridden with a settings node. On failure with skip_error enabled, the result becomes the error text instead of crashing the graph.

The honest take

Between this and the 3.1 Pro VLM: Pro judges, Flash captions. If your loop needs a strict art-director verdict, spend the extra credits on Pro with thinking on. If you're auto-captioning a dataset or doing batch sanity checks where determinism matters more than depth, this node is the better fit - it's cheaper, and the seed lets you lock a consistent answer across reruns. The detail combo the 3.1 vision nodes carry isn't here; the equivalent control is top_p/top_k and the instruction prompt, so set expectations in the system_instruction if you need fine-detail reading.

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. For reproducible captions in a batch loop, this is the vision node to leave wired in - the one you can rerun and trust to say the same thing twice.