Gemini Image Edit

Gemini Image Edit is the edit sibling of the pack's Gemini generate node: feed it an image plus a sentence, and Gemini 3 Pro Image rewrites the image to match. No mask input, no inpainting adapter - you describe the change and the model figures out the pixels. That's both the appeal and the honest limitation: it's pure instruction editing, closer to Qwen-Image-Edit's "give me a sentence, get a new image" behavior than to a controlled regional edit. If you need "change the shirt, leave everything else untouched," the pack's GPT Image Edit node with its mask is the better tool; this one is for the looser, more creative edits where you trust the model's judgment.

How it works

The node converts your image tensor to a JPEG (quality 95), base64-encodes it, and sends it as inline_data alongside your prompt to the same gemini-3-pro-image-preview:generateContent endpoint the generate node uses. The response's inline images get decoded back into an IMAGE tensor. One request, no polling - edit and done.

Inputs

• prompt - required. The edit instruction, in plain language ("make it sunset lighting," "turn the car red").
• image - required. Your input photo, as an IMAGE tensor.
• aspect_ratio - 1:1 (default), 9:16, 16:9, 4:3, 3:4.
• image_size - 1K (default) or 2K.

One output: IMAGE.

Installing

ComfyUI Manager (search "ComfyUI-MidjourneyHub") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jiaqianjing/ComfyUI-MidjourneyHub
# restart ComfyUI

Key goes in config.ini under [GEMINI_API] - the repo ships the author's key there, so swap it for your own before spending anything. Light dependencies, no model downloads.

Where people get burned

Because there's no mask, you can't protect a region - an edit that says "change X" will happily re-render the whole frame, and that's the thing users hit first. And Gemini image editing is a heavily filtered service (the KB's closed-source-models essay documents the strengthened safety filtering on Google's image stack), so don't expect the freedom of a local checkpoint. Add the pack's usual relay caveats - Chinese-language console, per-image billing, api_url subject to change without notice - and the honest summary is: a very capable model, accessed through an awkward middleman, best for loose creative edits rather than surgical ones.