diffvg-torch

A fast, pure-PyTorch reimplementation of diffvg. Differentiable and optimized for batched GPU workloads.

Key Features

• Pure PyTorch - No C++ compilation, no custom CUDA kernels
• Batched rendering - Efficient parallel rendering of multiple scenes
• Full gradient support - Backpropagation through the rendering pipeline

SVG Rendering

Cairo	diffvg-torch

Performance

Batched Rendering (render_batch_fast)

28x28 canvas, 1 path, 3 segments:

Batch Size	Total Time	Per Image
8	2ms	0.2ms
32	4ms	0.1ms
64	7ms	0.1ms

Backward Pass Speedup vs pydiffvg

28x28 canvas, 1 path, 3 segments:

Batch Size	diffvg-torch	pydiffvg	Speedup
8	5.2ms	16.2ms	3.1x
32	11.6ms	71.7ms	6.2x
64	23.1ms	137.9ms	6.0x

Demo: MNIST VAE

Vector graphics reconstruction using 1 bezier path with 3 segments per digit:

Installation

pip install torch numpy
pip install -e .

Quick Start

from diffvg_torch import render_batch_fast

# Render batched bezier paths
# control_points: [B, num_paths, num_segments, 4, 2]
# stroke_widths: [B, num_paths]
# alphas: [B, num_paths]
output = render_batch_fast(
    canvas_width=28,
    canvas_height=28,
    control_points=control_points,
    stroke_widths=stroke_widths,
    alphas=alphas,
    num_samples=4,
    use_fill=True,
)
# output: [B, 1, H, W] with gradients

SVG File Rendering

from diffvg_torch import render, svg_to_scene
import torch

canvas_w, canvas_h, shapes, shape_groups = svg_to_scene("input.svg")

output = render(
    canvas_width=256,
    canvas_height=256,
    shapes=shapes,
    shape_groups=shape_groups,
    num_samples_x=2,
    num_samples_y=2,
    background_color=torch.tensor([1.0, 1.0, 1.0, 1.0]),
    original_width=canvas_w,
    original_height=canvas_h,
)
# output: [H, W, 4] RGBA tensor

Supported SVG Features

• Shapes: <path>, <circle>, <ellipse>, <rect>
• Path commands: M, L, Q, C, Z (absolute and relative)
• Styling: fill, stroke, stroke-width, opacity
• Colors: Named colors, hex (#RGB, #RRGGBB), rgb(), rgba()

Citation

@software{diffvg_torch,
    title={diffvg-torch: Fast Differentiable Vector Graphics in Pure PyTorch},
    author={Carter McClellan},
    year={2025},
    url={https://github.com/CarterMcClellan/diffvg-torch}
}

License

MIT