Gemini Image Edit

"Make the sky stormy," "turn this product into a steampunk version," "swap the person's face with this reference." Where local inpainting needs a mask and a careful prompt, this node takes an image and a sentence and returns an edited version, because the Gemini image model understands both. It's the editing complement to the AIS_Gemini_ImageGen generator, and it uses the same closed model family - which is both the appeal and the price.

Required inputs: api_key, model (the same three image-capable models, default gemini-3.1-flash-image-preview), custom_model override, image (the image to edit, as an IMAGE tensor), and edit_instruction - the sentence describing what to change.

Then the pieces that make editing actually controllable:

• mask - a MASK input. Here's the honest mechanic, straight from the tooltip: Gemini doesn't do pixel-perfect masking. It uses the mask as visual guidance alongside your instruction, not as a hard boundary. White = edit, black = keep, and any MASK source works (SAM, Grounding DINO, segmentation, alpha channels, background removal). You get a loose regional edit, not surgical inpainting.
• reference_image / reference_image_2 / reference_image_3 - supply a face to swap in, a texture to match, a style to borrow. Reference-based editing is where this node is genuinely hard to beat locally.
• system_instruction - extra framing for the model.
• aspect_ratio and image_size - control the output dimensions (AUTO by default, which keeps things simple).

One output: image.

When to use it - and when not to

The sweet spot is the "in my head" edit: you can describe the change, so the model does the reasoning and the pixel-pushing together. Face swaps and character-consistency edits with a reference image are the strongest use case. The trap is treating it as free inpainting. If you need surgical, pixel-accurate edits to one region - a precise inpaint where only the marked pixels change - this is the wrong tool; use a local inpaint model. Gemini's edits can quietly rewrite more of the image than you asked for, which is fine for a creative pass and wrong for a preservation job.

Installing it

Part of the ComfyUI API Toolkit pack. Manager: search "API Toolkit". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-API-Toolkit
cd ComfyUI-API-Toolkit
pip install -r requirements.txt

Restart. Needs google-genai>=0.8.0.

Gotchas

• Every edit is a paid API call, and the pack re-runs on each queue - iterate cheaply by working at lower image_size first.
• Your source images leave the machine and Google's moderation applies. No celebrity faces, no copyrighted subjects, no workarounds; there are no weights to patch.
• The mask is guidance, not a contract. If the model edits outside the mask, lower your expectations for precision or switch to a local inpaint for that pass.