Gemini Generate Image

Gemini Generate Image is the pack's flagship cloud node: it calls Google's Gemini API through the official google-genai SDK and hands the result back as a normal ComfyUI IMAGE tensor - plus the model's text output, a thought trace, and the generated images as a list. One node, and you've got Google's image generation inside a local graph, no GPU required.

That last part is the whole pitch. Local sampling needs a serious GPU; Gemini runs in Google's cloud, so a laptop with a weak iGPU can still produce images. It's not a replacement for local diffusion - you're paying per image, subject to API quotas, and at the mercy of Google's safety and availability - but for "I need an image and I have no VRAM," it beats staring at your own hardware. The node also accepts up to five input images (image plus image1–image4), so it doubles as an image editor: send in a photo and prompt a change.

How it works

The node collects any connected input images, converts them to PIL, and sends a generation request to Gemini via the google-genai client. The default model_name is gemini-3-pro-image-preview - note that's the shipped default string, and image models rotate frequently, so if the call fails with a model error, check which image-capable models your key can actually access and set model_name accordingly. Results come back as a [N,H,W,C] tensor batch, with the response text and (if the model provides it) reasoning in the thought output.

Inputs that matter

• api_key - or the shared key file / GEMINI_API_KEY env var.
• vertex - flip on to route through Google Cloud Vertex AI instead of the consumer API.
• prompt - the image description; system_prompt for overall style guidance.
• aspect_ratio - auto, 1:1, 9:16, 16:9, 3:4, 4:3, 3:2, 2:3, 5:4, 4:5, 21:9.
• image_size - 1k, 2k, or 4k (default 2k). Bigger is sharper and pricier.
• seed - -1 for random, else fixed for reproducibility.
• thinking_budget / thinking_level - reasoning controls; 0 disables thinking, -1 leaves it unlimited, and thinking_level (OFF/AUTO/LOW/HIGH) is ignored unless the budget is 0.
• disable_safety_settings and safety_level - Google filters content by default; safety_level defaults to BLOCK_NONE and you can disable the settings entirely. Know your provider's ToS before leaning on these.
• base_url, proxy_url, timeout - endpoint override, proxy support, and timeout (0 = none).

Outputs

• image - first generated image as IMAGE; images - the full list; num_images - count.
• text - model response text; thought - reasoning trace if the model thinks.
• history - conversation history for chaining further calls.

Install & key setup

Part of ComfyUI-YogurtNodes; the Gemini nodes need google-genai, which the pack's requirements install:

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

Restart ComfyUI; the node is under "Yogurt Nodes" → LLM. Keys resolve in this order: the api_key field → api_key.json in custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/ ({"gemini": "YOUR_KEY"}) → the GEMINI_API_KEY env var. Get a key from Google AI Studio.

Common issues

• "Model not found" - the default gemini-3-pro-image-preview string may not match what your key can access. Set model_name to a currently available image model.
• "Auth failed" - check the three key sources in priority order.
• "Images are smaller/blockier than I hoped" - bump image_size to 4k; watch the cost.
• "Content filtered" - check safety_level / disable_safety_settings and your prompt wording.

This is the node I'd actually reach for among the pack's cloud options if I had to pick one - Gemini's image models are strong and the node surfaces the interesting extras (aspect ratios, thinking, editing) without burying them. Just keep a close eye on your API bill.