SvgIt

Raster → vector graphics, vectorizer.ai-style

Introduction

SvgIt turns raster images (PNG, JPG, …) into clean, compact SVG. It started as a fork of visioncortex/vtracer and grew a fully-owned tracing pipeline plus an optional ML layer (background removal, object segmentation, edge refinement), all wrapped in a small HTTP service with a live-preview UI.

The project is built in three levels:

• Level 1 — a thin axum service wrapping the VTracer core, with a parameter UI for live tuning.
• Level 2 — a dependency-free, fully-owned classical pipeline: LAB color quantization → region segmentation → contour tracing → RDP simplify → Schneider curve-fit → layering → minified SVG. Selectable as the Owned engine; produces exactly-N-color output.
• Level 3 — an ML layer (PyTorch → ONNX, run embedded in Rust via ort): salient-object background removal, FastSAM "segment everything", CNN edge/corner refinement, ×4 super-resolution for low-res inputs, and SCUNet denoise / JPEG de-blocking.

Quick start (the service)

cargo run -p svgit-service

Then open http://127.0.0.1:8080 and drag, paste, or upload an image — it converts live as you adjust parameters.

PORT=8090 cargo run -p svgit-service   # bind a different port

The first build downloads a prebuilt onnxruntime (via ort) and needs network access. Subsequent builds are fast. Add --release for much faster ML inference.

The VTracer and Owned engines work with no extra setup. The ML features need model weights — see below.

Engines

Engine	What it does
VTracer	The upstream vtracer core — stacked color clustering + curve tracing.
Owned	The dependency-free Level-2 pipeline; exact N-color flat output. Supports background removal, edge refinement, and gradient fills (detects smooth color ramps → <linearGradient>/<radialGradient>, merging quantization bands into one gradient).
Segment	FastSAM "segment everything" → layered SVG, one <g> per detected object.

ML layer & models

The weights are large and license-encumbered, so they're git-ignored and fetched on demand:

./scripts/fetch-models.sh             # all models, into ./models
SVGIT_MODEL_DIR=/path ./scripts/fetch-models.sh

Model	Size	Powers
u2netp.onnx	~4.6 MB	Remove background · Fast
isnet-general-use.onnx	~178 MB	Remove background · High (sharper fine detail)
lineart.onnx	~17 MB	Refine edges (owned engine contour snapping)
FastSAM-x.onnx	~289 MB	Segment engine
realesr-general-x4v3.onnx	~4.9 MB	Super-resolution (upscale low-res inputs 4×, any engine)
scunet_color_psnr.onnx	~91 MB	Denoise / JPEG de-block (clean artifacts before tracing, any engine)

Until a model is present, its feature returns a clear "model not found" error; every other feature still works. The core tracer needs no models at all.

Command-line app (vtracer)

The original VTracer CLI is still in the workspace:

cargo run -p vtracer -- --input input.jpg --output output.svg

It's also published independently on crates.io/vtracer (cargo install vtracer) and as a Python package (pip install vtracer).

OPTIONS:
        --colormode <color_mode>                 True color image `color` (default) or Binary image `bw`
    -p, --color_precision <color_precision>      Number of significant bits to use in an RGB channel
    -c, --corner_threshold <corner_threshold>    Minimum momentary angle (degree) to be considered a corner
    -f, --filter_speckle <filter_speckle>        Discard patches smaller than X px in size
    -g, --gradient_step <gradient_step>          Color difference between gradient layers
        --hierarchical <hierarchical>            `stacked` (default) or `cutout` (color mode only)
    -i, --input <input>                          Path to input raster image
    -m, --mode <mode>                            Curve fitting mode `pixel`, `polygon`, `spline`
    -o, --output <output>                        Path to output vector graphics
        --path_precision <path_precision>        Number of decimal places to use in path string
        --preset <preset>                        Use one of the preset configs `bw`, `poster`, `photo`
    -l, --segment_length <segment_length>        Subdivide-smooth until all segments are shorter than this
    -s, --splice_threshold <splice_threshold>    Minimum angle displacement (degree) to splice a spline

Workspace layout

Crate	Role
cmdapp (vtracer)	Upstream VTracer CLI.
webapp	VTracer WASM build for the browser.
service (svgit-service)	The axum HTTP service + live-preview UI.
pipeline (svgit-pipeline)	Owned, dependency-free classical tracer.
bgremove (svgit-bgremove)	ONNX background removal (u2netp / ISNet).
objseg (svgit-objseg)	FastSAM object segmentation.
edgenet (svgit-edgenet)	Line-art CNN edge map for contour refinement.
superres (svgit-superres)	ONNX ×4 super-resolution (realesr-general-x4v3).
denoise (svgit-denoise)	ONNX denoise / JPEG de-block (SCUNet).

Credits

SvgIt is built on VTracer by VisionCortex. See the tracing and clustering algorithm write-ups for the foundations it builds on.