Gemini Image Understand

Want a node that actually reads your images back to you - captions, questions about composition, structured data pulled out of a render - without downloading a single local model? This is that node. Gemini Image Understand takes up to five images and a prompt, ships them to Google's Gemini API, and hands you back text, a chat history you can chain, and even the model's thought.

It's part of the YogurtNodes pack (yogurt7771/ComfyUI-YogurtNodes), a grab-bag of ~150 utilities that includes a whole family of these LLM nodes - Gemini, OpenAI, Grok, Qwen, OpenRouter. The appeal is the same across all of them: your generation stays on the GPU, and the language/vision stuff happens in the cloud. Great for auto-captioning a batch before you LoRA-train, or for building a "describe what you see and re-prompt" loop.

How it works

Under the hood it's a thin wrapper around the google-genai client. Your IMAGE tensors get fed to whatever model you name in model_name (default gemini-2.5-flash), and the returned text becomes the text output. Because it's a real chat call, you also get a history output you can feed back into the node on the next run for multi-turn conversation - handy when you're iterating on a description.

The thought output is the sneaky-good part. With a thinking-capable model and thinking_budget set above 0, Gemini's chain-of-thought comes back on its own pin. I've seen people wire that into a second node to steer prompting. You don't have to use it; it's just there if you want it.

Inputs that matter

Most of the required fields are per-call knobs you can leave alone. The ones you'll actually touch:

• prompt - what you want said about the image. "Describe in detail" is a fine starting point; this is where you get specific.
• image, image1–image4 - up to five images. Wire in any IMAGE output. If you need more than five, feed them as a batch and loop, or string them through history.
• model_name - gemini-2.5-flash is the default and the sensible default. It's fast, cheap, and vision-capable.
• thinking_budget - 0 disables thinking (faster), -1 means don't limit it. Default 0 is the right call for quick captioning.

Then there's api_key - leave it blank and the node falls back to an api_key.json file in custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/ with {"gemini": "YOUR_API_KEY"}, then to the GEMINI_API_KEY environment variable. The README spells all three out. Outputs: text, history, thought.

Install

Same as every Yogurt node: ComfyUI Manager → search "YogurtNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes.git
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

The Gemini nodes need google-genai, which the pack's requirements install. Restart ComfyUI and look under Yogurt Nodes / LLM. No models, no checkpoints, no VRAM.

Troubleshooting

The README is blunt: if the API key isn't configured, "Gemini nodes will not work." Grab one at Google AI Studio. If your call fails, that's your first suspect.

A few things that bite people:

• vertex toggle - flips the client to Vertex AI mode (useful if you're on GCP and getting billed there instead). If you're not on Vertex, leave it off; a stray true gives you a confusing auth error.
• thinking_budget at 0 - this is the default, so if you were expecting deep reasoning and getting one-liners, that's why. Set it to -1 or a token budget.
• proxy_url - the tooltip accepts protocol://user:pass@addr:port including socks5/socks5h. If you're behind a proxy or VPN, this is where you set it.
• Big images, slow replies - there's no local encode; everything goes over the network. timeout defaults to 0 (never), so a stalled request can hang your queue. Setting a sane timeout in seconds is worth doing in production workflows.

If you just want a quick local sanity check that the key works before wiring a big workflow, prompt it on a test image with model_name left as-is - a one-liner back means you're good.