Skip to main content

​
Overview

comfyui-image-annotator is an interactive image annotation ComfyUI node contributed by community user luckdvr. Its core value: lowering the prompt-writing barrier. Instead of struggling to describe “replace the red button in the top-left corner,” you simply circle, click, or box the area you want to change and let the model see exactly where to edit. Designed to sit before the Nano Banana Pro node in a three-stage pipeline: “image → annotate → API.”
Project Info
  • 🔗 Source: github.com/luckdvr/comfyui-image-annotator
  • 📜 License: MIT
  • 👤 Author: luckdvr
  • ⭐ Community contribution — same author as Luck Nano Banana Pro
Recommended workflow: Image → Annotator → API node
LoadImage  ─►  ImageAnnotator  ─►  Luck Nano Banana Pro  ─►  SaveImage
               (mark the area to edit)      (edit per annotation + prompt)
Best fit: users who struggle with English prompts or precise “where to edit” descriptions. Let the mouse handle where, and a one-liner prompt handle what.

​
Core Features

Three annotation types

Points (⦿): single-click · Rectangles (▢): drag-to-draw · Polygons (⬡): multi-click with auto-close

Real-time rendering

Annotations render live on the canvas — WYSIWYG, no round-trip to preview

Zoom / pan / select

Built-in canvas controls let you annotate precisely even on large images

50-step undo

Up to 50 undo steps — experiment freely without fear

Dual outputs

Emits both the annotated image (for the model) and annotation JSON (for downstream parsing)

Customizable styling

Stroke color, stroke width, fill transparency, point size — all configurable

​
Supported APIYI Models

This node does not call any API on its own — it only annotates. The annotated image can feed any APIYI model that accepts image input. Best pairings:
ModelModel IDUseAPI Docs
Nano Banana Progemini-3-pro-image-previewRegion-targeted image editing / blendingView
Gemini / Qwen-VL familyvariousImage understanding, annotation-grounded VQAView

​
Node Details

​
Input

ParameterTypeRequiredDescription
imageIMAGEYesThe image to annotate

​
Outputs

OutputTypeDescription
annotated_imageIMAGEImage with rendered annotation marks (feed directly into downstream API node)
annotations_jsonSTRINGJSON string describing annotation positions/types (for nodes that need structured input)

​
Installation

1

Step 1: Clone into custom_nodes

cd ComfyUI/custom_nodes
git clone https://github.com/luckdvr/comfyui-image-annotator.git
2

Step 2: Restart ComfyUI

No extra dependencies — uses ComfyUI’s built-in environment. After restart, search ImageAnnotator in the node palette.
3

Step 3: Wire up the three-stage workflow

Connect the nodes in sequence:
LoadImage → ImageAnnotator → Luck Nano Banana Pro → SaveImage
Circle the area to edit on the ImageAnnotator canvas, then write a simple prompt in the downstream API node (e.g., “replace the marked area with a red sports car”) and run.

​
Usage Examples

​
Example 1: Local replacement (beginner friendly)

1

Load the source image

LoadImage — load a photo of your living room
2

Annotate the target region

ImageAnnotator — draw a rectangle around the empty space next to the sofa
3

Write a one-liner prompt

Luck Nano Banana Pro — prompt: “Put a green indoor plant in the marked area”
4

Run and save

The model adds a plant exactly in your box and leaves the rest untouched

​
Example 2: Multi-region precise edits

Use polygon to outline clothing + point for a hat location + rectangle for the background, then prompt: 
Change the clothing in polygon 1 to a black suit, add a hat at point 1,
replace the background in rectangle 1 with a sunset beach
The model will make distinct edits in each region.

​
Example 3: Visual Q&A / understanding

With Gemini / Qwen-VL: ask the model to explain “what is the object inside the rectangle and what’s its role in the scene.” Annotations give the model precise visual anchors.

​
FAQ

Traditional image editing requires you to describe both where and how — tough for non-native English speakers and AI beginners.With this node, location info comes from the mouse, so the prompt only needs to describe what to change. Descriptions that used to take 3-5 sentences become a single short phrase.
Yes — in a good way. Most multimodal models (Nano Banana, Gemini, Qwen-VL) recognize annotation symbols on the image as “user-indicated target regions” and follow instructions more precisely.If you’re worried about visual interference, lower stroke width and use semi-transparent fill in the node settings.
  1. Confirm the directory: ComfyUI/custom_nodes/comfyui-image-annotator
  2. Fully restart ComfyUI (frontend refresh is not enough)
  3. Check ComfyUI’s console for errors
Yes — just connect only the annotations_json output downstream. Useful when you want to pass coordinates into a custom script for post-processing.
Recommended pipeline: LoadImage → ImageAnnotator → Luck Nano Banana Pro
  • Wire ImageAnnotator.annotated_image into Luck Nano Banana Pro.image_01
  • Use natural-language prompts describing “what to do to the marked region”
  • If the first result is off, use Luck Nano Banana Pro’s retry_times or seed modes to re-run

Luck Nano Banana Pro Node

The author’s API-calling node — pairs perfectly with this one

Nano Banana Pro API

Full Nano Banana Pro capabilities

ComfyUI node collection

Browse all Nano Banana ComfyUI nodes

APIYI Console

Manage keys, usage, and channels