DiffVG is a differentiable rasterizer for 2D vector graphics. Tzu-Mao Li, Michal Lukáč, Michaël Gharbi, and Jonathan Ragan-Kelley published it as “Differentiable Vector Graphics Rasterization for Editing and Learning” in ACM Transactions on Graphics 39(6), Article 193, the proceedings of SIGGRAPH Asia 2020. The DOI is 10.1145/3414685.3417871. The project page lists Li and Ragan-Kelley at MIT CSAIL, and Lukáč and Gharbi at Adobe Research, which is the affiliation block of that paper.
The code is the GitHub repository BachiLi/diffvg, Apache License 2.0, in a LICENSE file at the repository root. The paper PDF carries the ACM copyright notice for the article. The license on the code and the copyright on the paper are different objects. On 6 October 2026 the GitHub API reported 1,282 stars, 214 forks, and 67 open issues. The releases list was empty. setup.py sets the package version to 0.0.1.
The bibtex key in the README is Li:2020:DVG. That key is this paper. A search for a newer product sold as DiffVG or DVG did not surface a separate 2025 or 2026 library under those names. The current reason to read the repository, if you are following SVG animation, is that newer methods still call it. LiveSVG (Levy et al., arXiv:2605.30174, 28 May 2026) names DiffVG as the renderer that fits SVG paths to a target video.
One frame, and the gradients
You give DiffVG shapes: paths, circles, and the other primitives in its scene, plus fills, strokes, and transforms. It rasterizes them to an image. Because the rasterizer is differentiable after pixel prefiltering, a loss on that image can send gradients back to curve parameters, colors, and related inputs.
The abstract describes two prefilters:
- An analytical prefilter. The authors say it is faster and can show artifacts such as conflation.
- Multisampling anti-aliasing. The authors say it remains efficient, renders a high-quality image, and computes unbiased gradients for each pixel with respect to curve parameters.
The project page keeps an errata note for the PDF. In Equation 6 the line integral’s measure should be along the boundary, dp(t), or the Jacobian of the curve should be included if the measure stays dt. In Equation 7 the integrand is their function g. The printed PDF uses f in that spot. Anyone checking a gradient derivation against the printed equations should read that note before treating a mismatch as a bug in the code.
The applications in the paper and on the project page are editing and learning tasks that compare a rasterization to a target image: an editor that optimizes under an image metric, painterly rendering by fitting random Bézier curves, a refinement pass on vectorization, seam carving applied through the rasterizer, and a VAE and a GAN trained with raster losses. Those demos produce vectors. The renderer itself still emits one image per parameter setting.
What “animation” means in this repository
The README is full of GIFs: a circle, an ellipse, a rectangle, a polygon, a curve, a path, a gradient, an outline, a transformed ellipse. Each GIF is an optimization recording, a shape moving because gradient steps are changing its parameters toward a target. The documented surface is those recordings plus the apps in the next section. A SMIL writer and a timeline are absent from the README.
LiveSVG’s preprint states the consequence directly: DiffVG produces a single image for a single set of SVG parameters. Papers that animate with it render one image per keyframe. LiveSVG’s own addition, described in the paired note, is to write the finished motion into animate tags inside one SVG file. That writing step is specified in the LiveSVG preprint.
A second name collision is worth keeping straight. LIVE, “Towards Layer-wise Image Vectorization” (Ma, Zhou, Xu, and colleagues, CVPR 2022), is a still-image vectorizer. Its repository says the implementation is based on the diffvg codebase, and its output is a static SVG. LiveSVG, the 2026 animation preprint, is a different project.
The code you can clone
The README’s install is a conda stack (PyTorch, numpy, scikit-image, cmake, ffmpeg) plus pip packages (svgwrite, svgpathtools, cssutils, numba, torch-tools, visdom) and python setup.py install. A second path uses Poetry and tells you to install Python 3.7. setup.py registers the distribution name diffvg at version 0.0.1, imports as pydiffvg when PyTorch is present, and also builds pydiffvg_tensorflow on non-Windows when TensorFlow is present. CUDA is turned on when torch.cuda.is_available() is true, unless the environment variable DIFFVG_CUDA overrides it.
The latest commit on master as of 6 October 2026 is 85802a71, 17 September 2024, Tzu-Mao Li merging pull request #86 (“fix: stroke-opacity”) in pydiffvg/parse_svg.py. The repository object’s pushed_at timestamp is 17 May 2025. That field updates when any ref is pushed. The master commit date is the September 2024 one.
The README’s examples live under apps/:
cd apps
python single_circle.py
python finite_difference_comp.py imgs/tiger.svg
python svg_brush.py
python painterly_rendering.py imgs/fallingwater.jpg --num_paths 2048 --max_width 4.0 --use_lpips_loss
python refine_svg.py imgs/flower.svg imgs/flower.jpg
python seam_carving.py imgs/hokusai.svgsingle_circle.py fits one circle to a target. svg_brush.py is the interactive editor. painterly_rendering.py throws random paths at an image. refine_svg.py adjusts an existing SVG toward a raster. seam_carving.py retargets through the rasterizer. The VAE script is apps/generative_models/mnist_vae.py. The GAN scripts are train_gan.py and eval_gan.py in the same folder. The path counts, widths, and iteration flags in those commands are the README’s examples.
Whether that install succeeds on a current PyTorch and a current compiler is open. The documented Poetry path still asks for Python 3.7, and there is no tagged release. Issue threads on the repository, including a Colab report closed in February 2026, are people negotiating toolchains. Treat those threads as field notes, and expect to pin a compiler and a PyTorch build before import pydiffvg works.
Where it sits next to LiveSVG
Use DiffVG when the problem is “change these vector parameters so the rasterized image matches a target.” Use LiveSVG’s paper when the problem is “make this existing SVG move, and give me one animated file.” LiveSVG’s fitting stage is the first kind of problem, repeated across 15 keyframes, with the motion stored in the SVG afterward.
The 2020 paper’s own numbers (timings of the two prefilters, reconstruction examples, VAE samples) are the authors’ demonstrations in that article. A 2026 hardware comparison would need a new run. The code you can audit is the Apache-2.0 repository. The animation file format you can open in a browser is the subject of the LiveSVG note.
Verdict
Clone the repository when a pixel loss needs to move Bézier paths. Start from the apps under apps/, and plan on the README’s older Python note before you assume a current toolchain. When the goal is one animated SVG, the file format and the fitting schedule live in the LiveSVG preprint, which calls this rasterizer per keyframe.
The Campfire
No commentsNobody has pulled up a log by this one yet. Be the first to say what you make of it.
Held for the desk. It appears after a look.