ComfyUI Primitive Mesh Generator

Image-to-vector art generator for ComfyUI - Convert photographs into artistic collages composed of geometric shapes using a greedy optimization algorithm.

This is a Python port of the Sketchbeast algorithm (itself a derivative of primitive.js by Ondřej Žára) adapted as a ComfyUI custom node.

Features

• Multiple shape types: Triangles, rectangles, rotated rectangles, ellipses, quadrilaterals
• Dual output: Returns both raster image (tensor) and vector SVG
• Reproducible: Seed-based random generation
• GPU Accelerated: Automatic CUDA/GPU acceleration for 1.3-1.5x speedup (30-55% faster)

Installation

Method 1: Direct Installation

1. Copy the comfyui_primitivemesh folder to your ComfyUI custom nodes directory:

   cp -r comfyui_primitivemesh /path/to/ComfyUI/custom_nodes/

2. Install dependencies:

   cd /path/to/ComfyUI/custom_nodes/comfyui_primitivemesh
   pip install -r requirements.txt

3. Restart ComfyUI

Method 2: Symlink (for development)

ln -s /path/to/comfyui_primitivemesh /path/to/ComfyUI/custom_nodes/comfyui_primitivemesh
cd /path/to/ComfyUI/custom_nodes/comfyui_primitivemesh
pip install -r requirements.txt

Usage

1. Add the "Create a Primitive Mesh" node to your workflow

2. Connect an image input

3. Configure parameters:

   • num_shapes: Number of geometric shapes to generate (7-500, default: 35)
   • shape_type: Type of shapes (mixed, rectangles, triangles, quadrilaterals, ellipses)
   • style: Visual style (crispy, dreamy, blurry) - controls shape transparency
   • seed: Random seed for reproducibility
   • candidates_per_shape: Number of candidate shapes evaluated per iteration (50-500, default: 150)
   • mutations_per_shape: Number of mutation attempts per shape (10-200, default: 30)

4. The node outputs:

   • image: Rasterized result as ComfyUI image tensor
   • svg_string: Complete SVG document as string

Note: compute_size is automatically set to 256 pixels for optimal performance

Algorithm Overview

The Sketchbeast algorithm works by:

1. Random Generation: Generate N random candidate shapes
2. Color Optimization: For each shape, calculate the optimal RGB color that minimizes visual difference from the target image
3. Selection: Choose the shape that best improves the approximation
4. Mutation: Refine the shape through M mutation iterations (small random modifications)
5. Application: Apply the optimized shape if it improves the overall result
6. Iteration: Repeat until reaching the target number of shapes

The distance metric is the root-mean-square (RMS) difference in RGB values across all pixels, normalized by image dimensions.

Output Usage

Raster Output

The image output is a standard ComfyUI tensor that can be:

• Saved using Save Image node
• Further processed with other nodes
• Previewed directly

SVG Output

The SVG string can be:

• Saved to disk using a text file node
• Further processed or styled
• Imported into vector graphics software (Illustrator, Inkscape, etc.)
• Scaled infinitely without quality loss