Vertex AI Generate Image

Here's the pitch: a node in your ComfyUI graph that generates images from Gemini's latest image model (gemini-3-pro-image-preview by default) and hands the result back as a real torch tensor that your local pipeline can keep processing - upscale it, VAE it, composite it, whatever. Local generation meets cloud generation on one canvas. That's genuinely useful for editing workflows where you want a model that isn't a checkpoint you have VRAM for. The catch, and it's a real one: this is Vertex AI, not a plain Gemini API key, which means Google Cloud service-account credentials and a project setup before it does anything.

How it works

The node calls Gemini through the google-genai SDK (that's the google-genai dependency in the pack's requirements) and returns the image as a torch.Tensor. Outputs:

• image - the generated image as an IMAGE tensor you can wire straight into a VAE decode or upscale path
• images - the full list, when the model returns more than one
• num_images - how many came back, so the rest of the graph can adapt
• text - Gemini's accompanying text (Gemini image models return a caption/description alongside the image)
• history - conversation history, for chaining edits
• thought - the model's reasoning trace when thinking is enabled

Optional image through image4 inputs turn it into an editing node: feed in an image and the prompt describes the edit. That's the real strength here - "change the lighting" against a rendered frame, using a frontier model, then feed the result back into your local graph.

The inputs that actually matter

• credentials - the Vertex AI credentials JSON. This is the wall most people hit. It's a Google Cloud service-account key, not an AI Studio API key.
• project_id and location - your GCP project and the Vertex AI region it lives in. All three are required and there's no default.
• prompt (multiline) - what you want. system_prompt sets the overall style.
• image_size (default 2k) and aspect_ratio (default auto) - Gemini-native output settings. 2k is the big one; drop it if you want faster returns.
• seed (default -1) - random seed; set a fixed value for reproducibility, which matters when you're using this as a node in a pipeline you iterate on.

The defaults are worth noticing: safety_level ships as BLOCK_NONE and disable_safety_settings defaults to false, so out of the box it runs with minimal filtering. Google's platform still applies its own content rules on the API side, but the node isn't adding an extra censorship layer on top. If you want tighter filtering, change safety_level.

The honest cost

Vertex AI is a cloud API with a billing account behind it. Every generation costs money, and you're putting your GCP credentials inside your ComfyUI graph - a real security consideration in a tool that, as the ecosystem's history with malicious nodes shows, executes whatever a custom node imports. Don't share a workflow containing your credentials JSON with strangers, and don't leave a public ComfyUI instance reachable while this node is configured. It's not a reason to avoid the node; it's a reason to be deliberate about it.

Installing and setting up

Install the pack once via ComfyUI Manager (search "ComfyUI-YogurtNodes") or:

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

Then you need a Google Cloud project with Vertex AI enabled, a service-account key (JSON) with permission to call it, and the project ID + region. Paste the JSON into credentials, fill project_id and location. If you'd rather use a plain Gemini API key from AI Studio, that's what the pack's separate Gemini Generate Image node is for - this one is specifically the Vertex path, and the tooltips make that split explicit.

Common issues

Every failure here is a config failure. "401 / permission denied" - the service account lacks Vertex AI access; add the role in GCP. "Invalid region" - location has to be a region where the model is available (us-central1 is the usual starting point, though check current availability). Empty credentials - the node has nothing to authenticate with. And if you get timeouts, that's what timeout (default 0 = no limit) and retry_count exist for. Start with the smallest image_size, one seed, and confirm you get one image back before building anything ambitious on top of it.