SF VertexAI Nano Banana Pro Edit

Nano Banana Pro is the closest thing closed-source image gen has to a flagship right now - Gemini 3 Pro Image, native 4K output, and a "thinking mode" where the model reasons about your prompt before it draws anything. It's also, of course, impossible to run locally. SF VertexAI Nano Banana Pro Edit is the node that drops that model into your ComfyUI graph: feed it one to many reference images, tell it what to change, and get back an edited image plus the model's thinking.

The practical magic of this node is the multi-image editing. Want a character kept consistent while the outfit, background, and lighting all change? Give it a few reference shots and one instruction sentence. That's the workflow this was built for - and it's the same thing that makes the node fiddly, so read the inputs before you click.

It's part of the comfyui-stillfront pack, under Stillfront/VertexAI.

How it works

The node calls Gemini's image-edit model through Vertex AI (google-genai with vertexai=True), sending your images and the edit_instruction together. Auth is standard Google Application Default Credentials - gcloud auth application-default login, project with the right APIs enabled. Note the location default here is global, not us-central1 - the tooltip spells out why: use global for Gemini/Nano Banana models.

One thing to know upfront: every image it produces carries Google's invisible SynthID watermark, and Google's content policy is strict. Nano Banana is powerful but, as the community puts it, "unimaginative" under heavy filtering. This is a Google-rules node, not an uncensored one.

The inputs that matter

• edit_instruction - the whole game. Be specific: "Make it look like a cartoon" or "Change the background to a beach" are the tooltip's examples, and that style of direct instruction is what works.
• model - two choices: plain Nano Banana (gemini-2.5-flash-image, faster, 1K only) and Nano Banana Pro (gemini-3-pro-image-preview, best quality, up to 4K). Default is Pro.
• image_size - 1K/2K/4K. Nano Banana only supports 1K; Pro handles all three.
• aspect_ratio - ten options, from 1:1 up to 21:9.
• seed - default 0 means random; set it for reproducible edits.
• inputcount - this is the fiddly one. It defaults to 4 and can go to 14, but the tooltip warns you must click the "Update inputs" button after changing it for new image slots to appear. There's no need to fill all slots - image_1 is required, image_2 and image_3 are the visible optional ones, and the rest exist for when you crank inputcount up.

Outputs are three: edited_image (IMAGE), plus text_response and thoughts (both STRING) - the model's answer text and its reasoning, respectively. The thoughts output is a genuinely fun debugging tool: when an edit comes out wrong, read what it was thinking before you rewrite the prompt.

Installing it

Pack install, as always:

cd ComfyUI/custom_nodes
git clone https://github.com/razvanmatei-sf/comfyui-stillfront
cd comfyui-stillfront
pip install -r requirements.txt

Restart, then set up Google Cloud: enable the API, gcloud auth application-default login, and either fill in project_id or export GOOGLE_CLOUD_PROJECT.

Troubleshooting

• Changed inputcount and no new slots appeared. Forgot the "Update inputs" button. Click it.
• Edits keep getting refused or come back bland. That's Google's safety net, and it's aggressive on people, IP, and celebrity content. Rephrase the instruction; don't fight it.
• Auth failures. Same ADC story as every Vertex node in this pack.
• Cost creep. Nano Banana Pro at 2K/4K with multiple reference images adds up fast - the community estimates roughly $0.04–$0.24 per image depending on resolution. The 1K Nano Banana is the cheap exploratory option before you commit to the expensive render.