FL Vertex Gemini 2.5 Flash Image

Fill-Nodes has two different ways into Google's Gemini 2.5 Flash Image model, and it's worth knowing which one you're looking at. FL_GeminiImageEditor is the simpler consumer-API route. FL_VertexGemini25FlashImage is the other one: it authenticates through Google Cloud Vertex AI using a service account, not a plain Gemini API key. That's a meaningfully bigger setup lift - you need a GCP project with Vertex AI enabled and billing configured, plus a service account with the right role - but it's the route that fits if your image generation already has to live inside a GCP environment for compliance, quota, or billing-consolidation reasons.

Fill-Nodes as a whole (filliptm's sprawling grab-bag pack) leans heavily on third-party paid APIs across its AI section - Fal, OpenAI, Runway, and Google in a couple of different flavors. This is the "enterprise-adjacent" one of that group.

How it works

You give it a prompt, a service account credential, and up to three reference images. It authenticates against Vertex AI with the service account, sends the request to the Gemini 2.5 Flash Image endpoint, and runs batch_count generations in parallel - so a single node call can return multiple variations from one prompt rather than requiring you to trigger the graph repeatedly.

The inputs and outputs that matter

• prompt - your generation prompt.
• service_account_json - the GCP service account credential. This is the field that trips people up (see below): it needs your service account's actual JSON key content or a path to it, and that service account needs Vertex AI access in a project with billing enabled.
• batch_count (default 1, 1–8) - how many images to generate in parallel from the same prompt.
• temperature (default 1, 0–2) - generation variability; lower for more literal/consistent results, higher for more varied ones.
• seed (default 0, capped at 666666 rather than a typical 32-bit max - a genuinely unusual ceiling worth just knowing about rather than wondering if you mistyped a bigger number).
• Optional reference_image_1/_2/_3 - up to three reference images the model can draw on.

Outputs are images (the generated batch) and API_Response - a detailed log string worth routing into FL_ShowText while you're setting this up, since Vertex AI auth failures are far easier to diagnose from the actual response text than from a bare ComfyUI error.

How to install it

ComfyUI Manager: search "Fill-Nodes", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes
pip install -r ComfyUI_Fill-Nodes/requirements.txt

This pulls in Google Cloud client libraries alongside everything else Fill-Nodes needs - one more reason this is a genuinely large install for what looks, from the outside, like a single image-generation node.

Common issues & troubleshooting

Auth failures right out of the gate. By far the most common issue with Vertex-AI-backed nodes generally: the service account is missing the Vertex AI User role, the GCP project doesn't have the Vertex AI API enabled, or billing isn't set up on that project. Read API_Response - Google's error messages are usually specific about which of these it is.

Confusing this with the simpler Gemini nodes in the same pack. If you just want a quick Gemini API key and don't need GCP infrastructure, FL_GeminiImageEditor (Google's direct Gemini API) is the far less involved route - reach for this Vertex node specifically when you need it inside GCP, not by default.

Seed values above 666666 get rejected or clamped. That's the node's actual max - not a typo on your part. Stay under it.