Uncut NBP (Vertex AI)

This is a ComfyUI node that calls Google's Nano Banana Pro (Gemini 3 Pro Image) through Vertex AI - with the safety filters switched off and no API key sitting in your workflow. If that already sounds like something you can't just fire up on your home rig, you're right, and the README is refreshingly honest about it. Before you get excited, read the one line that decides whether this node is for you: the key to Vertex AI "is supplied through internal channels." In plain English, this is a work-internal tool. If you're not inside whatever company runs the Uncut token relay, this node will fail on its very first step.

Why it exists. Nano Banana is the closed image model everyone keeps pulling into open workflows - you generate frames on it, then animate them with Wan or another local video model, the canonical hybrid of 2026. But Google's API path has two annoyances for an organization: you need a Google Cloud project with Vertex AI and a billing account, and the filters are aggressive enough that a production pipeline trips them constantly. This pack is the "Uncut" answer: an org that has IAM permission to set Google's safety thresholds to OFF, and a local relay server that hands out short-lived Vertex access tokens so nobody's key ends up in a workflow JSON that gets shared around.

How it works. The node talks to Google's google-genai SDK with vertexai=True, your GCP project_id and location. The interesting part is server_ip: instead of a key on the node, it does a GET http://<server_ip>/get_vertex_token and uses the returned access_token as its credentials. Everything else is standard Nano Banana generation - system_instruction, an aspect ratio and resolution baked into the image config, a seed, and the model choice. It even ships a decoder that maps Google's error support codes to plain words ("child", "celebrity", "sexual", "dangerous content") so when the filter does trip, you know exactly why instead of squinting at hex.

Inputs that matter. Most of these you'll leave alone:

• model - gemini-3-pro-image-preview (Nano Banana Pro, the 4K flagship) or gemini-3.1-flash-image-preview (Nano Banana 2, the faster hybrid). Default is Pro.
• server_ip - your org's token relay, default 127.0.0.1:9191. This is the thing that decides if the node works at all.
• project_id / location - your GCP project and region (default global).
• aspect_ratio and resolution - seven ratios from 1:1 to 16:9, and 1K / 2K / 4K. Note 4K is the Pro model's native playground, but it's also where your bill climbs.
• reference_images - one optional IMAGE tensor. It's converted to PIL and sent alongside the prompt, which is how you get the subject-consistency editing Nano Banana is famous for.
• seed - deterministic output, default 0.

The single output, generated_image, is a normal ComfyUI IMAGE, so it wires straight into a save node, an upscaler, or a local video model for that hybrid workflow.

Install. ComfyUI Manager → Custom Nodes Manager → "Install via Git URL", paste the URL, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/This-is-Uncut/Custom-Node-ComfyUI-NBP-Uncut-Vertex
cd Custom-Node-ComfyUI-NBP-Uncut-Vertex
pip install -r requirements.txt

Dependencies are light: google-genai>=1.55.0, pillow, numpy, requests. No model files - everything runs on Google's side. The pack has an install.py that pip-installs requirements automatically, but if ComfyUI errors on startup or you're not getting high-res output, run that pip line manually and bump google-genai - an outdated SDK is the usual culprit.

Where people get burned. First, the access problem: without the relay server reachable, you get Failed to fetch Vertex token from http://...:9191/get_vertex_token and the node is dead on arrival - that's not a bug you can fix from the node. The README also warns the first generation may need a retry while it fetches the key. And expect occasional 429 quota errors; the node retries with exponential backoff for you, but a busy team can still throttle a batch. If you're a hobbyist who just wants Nano Banana in Comfy, skip this one - the official ComfyUI API Nodes or any Gemini-key pack will get you the same model without needing a private server you'll never have access to.