Vertex Image Understand

Every serious workflow eventually wants a critic in the loop. You render, you look, you tweak, you render again - and a vision model can do that looking step automatically. YogurtVertexAIImageUnderstand is the pack's image-to-text node: feed it an image (up to five), ask it something, and it returns a text description or judgment you can route anywhere in the graph. Caption a render, check that a face came out right, extract what's actually in a frame - the description on the box is "Understand images using Vertex AI," and that's the job.

How it works

It sends your images to Gemini (default model gemini-2.5-flash, a capable and cheap vision pick) through the google-genai SDK, along with your system_prompt and prompt. Outputs:

• text - the model's answer as a STRING, ready for a display node or a conditional
• history - the conversation, for follow-up questions
• thought - reasoning trace when thinking is on

Optional image through image4 inputs take the pictures. The auth story is the same as the pack's other Vertex nodes: a vertex toggle decides between a plain Gemini API key (api_key, AI Studio) and full Vertex AI (credentials JSON + project_id + location). Start with vertex off and an AI Studio key - for describing images you don't need Google Cloud.

The patterns people actually build

The three that earn their keep:

• Captioning / tagging. Feed a render, prompt "list the main subjects, lighting, style, and any flaws," and feed the caption back into a prompt-schedule node for style consistency across a batch.
• Quality gate. Ask "is this image sharp and well-composed?" and route the answer through a YogurtSwitch to decide whether to rerun or accept. This is where a workflow starts to feel autonomous.
• Reference description. Describe a reference image so a text-only pipeline can recreate it - a way to turn "this look" into words.

Because text is a plain STRING, the output drops into YogurtStringToValue, YogurtSwitch, or any text socket. That's the nice part of this node's design: it produces the one type everything else in ComfyUI already speaks.

The inputs that matter

• prompt (multiline) - what you want to know about the image. Be specific; "describe this" gets you less than "list the objects, lighting, and any visible artifacts."
• system_prompt - sets the mode ("you are a strict art critic" beats no system prompt for quality-gate work).
• temperature (default 1) - for judgment tasks, lower is steadier.
• thinking_budget - a small non-zero budget improves visual reasoning noticeably. Worth experimenting with.

Installing it

Part of the Yogurt Nodes pack (YogurtNodes/LLM), one install for everything:

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

Or ComfyUI Manager, search "ComfyUI-YogurtNodes," install, restart. API key via GEMINI_API_KEY env var or api_key.json in custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/ - the README has the file format.

Common issues

Same config pitfalls as the sibling Vertex nodes: wrong vertex toggle, empty credentials, a location where the model isn't served. One this node adds: it's sending your images to Google's servers. If you're generating things you'd rather not upload, don't wire this into that pipeline - use a local vision node instead. And remember it sees a downscaled/encoded version of your image, so it can describe composition and obvious flaws, but don't expect it to read a 4K frame's fine text. Keep the prompt realistic about what a vision API can do.