Gemini视觉丨API

If you want Google's vision models inside ComfyUI without messing with the official node ecosystem, GeminiVisionAPI is the QING pack's door in. It sends an image and a question to Google's Gemini API and returns the analysis as text. The 2.5-series models get a reasoning_effort knob - none/low/medium/high - which is the thinking-depth control for the flash and pro models, and the one thing that separates this from a dumb image-to-text call. Feed it a frame from your video workflow and ask "describe the composition," and the answer comes back ready to steer your next prompt.

It's part of ComfyUI-QING's API category, platform fixed to Google AI Studio (no dropdown to confuse). The model field lists six Gemini options from gemini-2.0-flash up through gemini-2.5-pro, with gemini-2.5-flash-lite as the default - the lightweight one, which is the sensible default for caption-and-move-on jobs.

How it works. The image and your text_input go to Google's Gemini endpoint using the key you configured in the QING settings panel. The framework handles the multi-turn history (default 6 rounds) and token limits. The interesting bit is reasoning_effort: for the 2.5 series this controls how much the model "thinks" before answering - none is fast and direct, high burns tokens but gets you deeper visual reasoning. The tooltip is explicit that this is a 2.5-series-only feature, so it's a no-op on gemini-2.0-flash.

The inputs that matter.

• image - what you're analyzing.
• text_input - your question (default "describe this image").
• model - six choices; gemini-2.5-flash-lite default.
• max_tokens (default 4096) and history (default 6 rounds).
• Optional: temperature (default 0.4 - low, for consistent analysis), top_p, image_quality, reasoning_effort (default none), clear_history.

Outputs: analysis_result, conversation_info, total_tokens.

How to install. ComfyUI-QING is a one-pack install - search "ComfyUI-QING" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

Set your Google AI Studio key in Settings → QING → API Configuration and restart. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. The reasoning-effort gotcha: it only does anything on 2.5-series models, so cranking it to high on gemini-2.0-flash silently changes nothing while you wait longer for nothing. Second: Gemini 2.5 models can be slow, and "high" reasoning plus a big image plus max_tokens 32768 is a recipe for a timeout - drop the reasoning effort or the token cap if calls are timing out. Third, your key lives in the QING settings panel, so a fresh install that never opened Settings → QING will fail auth until you add it. And as with every API node here: your image and prompt leave the machine, and Google's content policy applies to whatever you ask about.