OpenAI Image Understand

You've generated an image and you need to know what's actually in it - for a caption, a quality review, a "does this match my prompt" check, or just a second opinion. YogurtOpenAIImageUnderstand is the vision-model node that answers: feed it an image, ask it a question, get a STRING back.

The setup mirrors the pack's other OpenAI nodes:

• api_key - inline, or blank to fall back to api_key.json / OPENAI_API_KEY.
• model_name - defaults to gpt-5.1, called as a vision model.
• system_prompt and prompt - the framing and your question ("Describe this image", "What's wrong with this composition?").
• temperature, top_p, max_tokens, frequency_penalty, presence_penalty - the standard sampling dials.
• retry_count, timeout, proxy_url - resilience settings.
• chat_template - the default <-system->/<-user-> template, restructurable.

Optional inputs are the star here: image plus image1–image4 let you feed up to five images for comparison or multi-image questions. history chains turns, extra adds raw JSON to the request, and image_send_mode picks the wire format - openai (standard image_url objects) or base64 (for x.ai-style compatibility). Outputs: text, history, and payload (the raw response).

The workflows this unlocks

The text output is a first-class STRING, so it plugs straight into logic. Three patterns that actually get built:

1. Captioning loops. Generate → caption → feed the caption into your next prompt. Self-improving iterations without you in the middle.
2. Quality gates. Ask "does this image match the prompt: <prompt>?" and branch on the answer. Auto-reroll until the model says yes.
3. Reverse engineering. Drop in a reference image, ask for a prompt that would recreate it, wire the reply into a text2img node. It's not perfect prompt extraction, but it's a shockingly good starting point.

Multi-image input is the sleeper feature - "which of these two renders is closer to the reference?" is a real question you can ask in one call.

Install and keys

Ships in ComfyUI-YogurtNodes. ComfyUI Manager → search ComfyUI-YogurtNodes → install, or:

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

Restart ComfyUI, find it under "Yogurt Nodes/LLM". Needs the openai package and a valid API key; nothing to download.

Two caveats. The default model_name is a text-model default - confirm it's actually a vision-capable model before you count on image understanding, or you'll get an "unsupported image" error for your trouble. And remember the answer is only as good as the question: "describe this" gives you a paragraph of fluff, "list exactly three visible objects and their positions" gives you something a workflow can act on.