Imagen Edit (Vertex AI)

Here's the thing about instruction-editing models like Qwen or Kontext: they're brilliant at "change the whole vibe," but they touch every pixel, and that drift compounds across edits. Imagen Edit is the other school - you give it a mask and it only touches what the mask covers. The unmasked pixels come back untouched, which is exactly what you want for surgical fixes. In a world that mostly abandoned masks, this node is a reminder that they still own that job.

What it is

An image/edit node that sends an image, a mask, and a prompt to Google's Imagen editing model. The edit_mode dropdown is the headline feature and covers the four real use cases:

• EDIT_MODE_INPAINT_INSERTION - add something into the masked area (default).
• EDIT_MODE_INPAINT_REMOVAL - delete what's in the mask, fill it in.
• EDIT_MODE_OUTPAINT - extend the image beyond its edges.
• EDIT_MODE_BGSWAP - swap the background.

The mask is user-provided, which makes this the natural partner for the Gemini Segmentation node in this pack: describe what to mask, get a MASK, feed it straight into mask. Two nodes, one coherent edit pipeline.

How it works

Under the hood it calls Vertex AI's edit_image with your image and mask as reference images (mask uses MASK_MODE_USER_PROVIDED with a mask_dilation buffer). The model is hardcoded to imagen-3.0-capability-001 - there's no model input on this node, so don't go looking for one. You get back edited_images, one or more, decoded from the response.

The inputs that matter

The credential trio, same as the whole pack - project_id, location, service_account, and yes, service_account wants the JSON contents pasted in, not a path.

• image + mask - your source and the edit region. The mask defines where; the prompt defines what.
• prompt - the edit instruction. Defaults to the uninspired "Edit this image"; say what you actually want.
• edit_mode - pick your operation (see above).
• mask_dilation (default 0.03) - expands the edit zone slightly past the mask edge. Bump it if you see a visible seam where the edit boundary meets untouched pixels.
• base_steps (10–100, default 50) and guidance_scale - the quality/speed dials.
• number_of_images - 1 to 4 candidates per run.
• negative_prompt - optional, and useful for "no text" type constraints.

Outputs

edited_images - an IMAGE output (stacked if you asked for multiple). Preview and compare against the source; for surgical edits, the comparison is the whole point.

Installing

cd ComfyUI/custom_nodes/
git clone https://github.com/Aryan185/ComfyUI-VertexAPI.git
cd ComfyUI-VertexAPI
pip install -r requirements.txt

Restart ComfyUI, or install ComfyUI-VertexAPI via ComfyUI Manager.

Where people get burned

The mask quality is everything. A sloppy mask gives you a sloppy edit, so make the mask carefully or let Gemini Segmentation make it for you. Expect the same Google safety filtering as the rest of the lineup - this won't do things your local inpaint shrugs at. And the "only touches the mask" guarantee is real but not absolute: at high base_steps or with a badly dilated mask, the edited region can bleed. If you see seams, dial mask_dilation up a touch; if you see the edit creep past its bounds, dial it down. Small numbers, big difference.