OpenAI Generate Image

The awkward truth about cloud image models in ComfyUI is that they output files, and ComfyUI wants tensors. YogurtOpenAIGenerateImage closes that gap: it calls the OpenAI image API and hands you back a proper IMAGE tensor - no downloading, no re-encoding, no PNG-loading node to wedge in between.

The inputs that actually matter on day one:

• api_key - your OpenAI key. Leave blank and it falls back to the pack's api_key.json file or the OPENAI_API_KEY env var.
• prompt - the main prompt.
• system_prompt - a system-level instruction that biases overall style.
• model_name - defaults to gpt-image-1.
• size, quality, style - resolution, quality level, and style (vivid/natural). Note style is dall-e-3 only, and n (number of images) is dall-e-2 only - the tooltips say so because the API mixes generations of models.

Then there's the whole second layer: base_url (blank = official API, but set it for Azure or any OpenAI-compatible server), chat_template (the default template slots {{system_instruction}} and {{prompt}} into a structured message), seed, aspect_ratio (for nano banana models), proxy_url, api_type (auto picks between the Responses and Images API based on model), timeout, and retry_count.

The optional inputs are where it gets interesting: image plus image1–image4 let you do image editing - feed a base image and prompt a change. There's also history for multi-turn, extra for raw JSON request params, and image_send_mode (upload for official, base64 for x.ai-style compatibility).

Outputs: image (single IMAGE), images (the batch as a list), num_images, text (the model's text response), and history.

What you can do with it

The tensor output means the result flows straight into standard ComfyUI nodes - VAE-free because it's already an image, ready for upscaling, compositing, comparison, or a detail pass. Pair it with the pack's YogurtSaveImageBridge to write files, or use the text output to capture the model's explanation of what it generated. The image/image1-4 edit inputs make it a poor-man's photo editor inside your graph: fix a face, restyle a product shot, iterate on a composition, all with one node.

Install and keys

Ships in ComfyUI-YogurtNodes - a quiet, single-maintainer pack with basically no community footprint, so the README is your documentation. 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". The openai Python package is the only real dependency here.

Three things get people:

1. Key not picked up. The priority is inline api_key field > api_key.json in custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/ > OPENAI_API_KEY env var. If you filled the field, it wins - that's usually fine, just know it can be a file instead.
2. DALL-E-3-only params. style and n do nothing on gpt-image-1. Read the tooltips.
3. Cost. This is a metered API. retry_count and a big n multiply spend; keep n at 1 and watch your usage panel.