Gemini (Soze)

The headline trick of this node is right in the defaults: drop it in a workflow, leave operation_mode on generate_images and model_name on gemini-3-pro-image-preview, and you're calling Gemini's flagship image model - the one Google rebranded as Nano Banana Pro - from your ComfyUI graph. No GPU, no checkpoint download, nothing local. Just a Google API key and an internet connection.

It's a single-node pack: "Gemini Node", displayed as Gemini (Soze) under the "Soze Nodes" category. SozeInc ported it from the long-running ComfyUI-IF_Gemini to talk to the Gemini 3.0 API, trimmed the Open Router support, and shared it back. The README is upfront that this started as a personal-use port - it's new, lightly tested, and does exactly one job. That job is genuinely handy.

What you'd use it for

Three modes, one node:

• generate_images - text-to-image (or reference-image-to-image) with Gemini's native image generation. Wire the image output to a Save Image node and you're done. It also means the output is a real ComfyUI tensor, so you can run it through an upscaler or a detailer the way you would any other generated frame.
• analysis - feed an image in, get a text description out. The default prompt is literally an image-to-prompt describer ("describe this as a film director would"). Pipe an image in, grab the text output, feed it to a KSampler - that's the classic IF_Gemini workflow, and it's still the best reason to install this.
• generate_text - plain LLM chat/response mode for a system prompt plus your instruction.

How it works

Under the hood it's the google-genai SDK doing a network call on every queue. The node resolves your API key in order: GEMINI_API_KEY environment variable, then your .zshrc/.bashrc exports, then .env files in a few locations, then whatever you paste into external_api_key. There's a Verify API Key button on the node that hits the pack's /gemini/check_api_key route so you don't discover a bad key mid-run.

The inputs that actually matter

• operation_mode - analysis, generate_text, or generate_images. Start here.
• prompt - your instruction. Multiline, and it's pre-filled with the image-description prompt, which trips people up in generate_images mode; replace it.
• model_name - the six-model list, defaulting to gemini-3-pro-image-preview. If you pick a text-only model while in generate_images mode, the code silently swaps you back to the image model - helpful, but don't be surprised when your choice is ignored.
• thinking_level - minimal/low/medium/high controls how much Gemini "thinks" before answering (Gemini 3 only). high is slow and pricier; leave it on minimal unless you want the reasoning depth.
• images (optional) - reference images for style-based generation or analysis. max_images caps how many get sent (default 6).
• aspect_ratio, media_resolution, batch_count, seed - aspect presets, 1K/2K/4K output, up to 20 images per run, and a seed for repeatability.

Outputs are text (STRING) and image (IMAGE). In analysis/text modes, image is just a placeholder.

Installing it

ComfyUI Manager search "comfyui_soze_gemini" (it's on the Comfy Registry under publisher sozeinc), or:

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

Restart ComfyUI. The dependency list is light - google-genai, google-generativeai, python-dotenv, plus torch/Pillow you already have. No model files. The real prerequisite is a key from aistudio.google.com/apikey.

Where people get burned

• It costs real money. Nano Banana Pro is billed per image - roughly $0.04 up to $0.24 depending on resolution. batch_count at 4K can eat a couple of dollars in one run. Keep an eye on the meter.
• The censorship is server-side. The code sets BLOCK_NONE on safety categories, but Google still enforces its IMAGE_SAFETY policy on the API; "Powerful, unimaginative, massive hidden biasing" is the community's summary. If a run returns a safety error, reword the prompt - the README says so too. It's not the node being broken.
• Every image carries an invisible SynthID watermark. Fine for personal work, worth knowing before you ship something commercial.
• Quota and key errors return as text output. If your text output suddenly says "ERROR: API quota exceeded" or "No Gemini API key found", that's the node talking to you, not a crash. The Verify button sorts out which.
• The pack is a young personal port, and exactly one node ships - the prompt-manager helpers that shipped in IF_Gemini are commented out here. Don't expect polish, expect "works".

The name is a mouthful and the pack is barely a week old, but if you want Nano Banana generations or an image-to-prompt step inside ComfyUI without standing up a separate service, this is the node to reach for.