Inkvec Trace (raster to SVG)

"Flux is terrific for quickly generating logos" is a true sentence right up until somebody asks for the file. On the thread where that was posted, the comment that stuck was a guy who'd generated a logo he wanted to use for his company, then realised it needed to be a vector - "then it was on to illustrator." That's the gap Inkvec Trace closes: Recraft is the one commercial generator that emits SVG natively, so everywhere else, ComfyUI included, you generate a raster and convert afterwards.

What it actually is

Inkvec Trace is a thin ComfyUI wrapper around Inkvec, LogoLabs' raster-to-SVG tracer. Not a diffusion model doing something that looks like tracing - a purpose-built vectoriser for flat artwork: logos, icons, diagrams, stickers, UI graphics. Photographs are outside its design scope, and one test run shows you why: banding, hundreds of near-identical paths, a file bigger than the PNG.

Under the hood it shells out to the native inkvec Rust binary, so the tracer's behaviour is the binary's behaviour.

How it works

The tracer reduces your image to a palette of at most colors inks, merging any two colours closer than merge in OKLab - the "these are the same colour" knob, and why a JPEG-ish gradient collapses into a handful of flat fills instead of a mess. Then it fits geometry to each ink region.

precision is where the real dial is. It sets the MDL cost of a coordinate, lambda = ln(extent / precision): smaller keeps more detail with more points. Despite the name it does not control output decimals - coordinates are always written with two. min_area throws away specks below that many square pixels, your despeckle for noisy input.

Defaults are the CLI's own, and the node passes every numeric widget explicitly, so what you see is what runs.

Inputs that matter

• image - the raster. Anything over max_dim (2048) on its long side is traced at 2048 while the SVG is still written at the original size. Nice default, but a 6K scan is then fitted from a downscaled view; raise max_dim if hairlines vanish.
• mask - transparency, LoadImage's convention (1 = transparent). Connect LoadImage's MASK and holes survive into the SVG. A segmentation mask is the other way round: run Invert Mask first.
• lossy - noise-aware intake for compressed input. The trap: the CLI's auto sniffs the file type, but ComfyUI hands the node decoded pixels, written as PNG, so auto behaves like a PNG here. Set it on for JPEG/WebP sources and anything from Inkvec Denoise.

native_alpha (on, inkvec 0.1.4+) traces transparency properly - inks carry opacity, holes stay holes, so white artwork on transparency survives. Leave it on. no_background drops the face covering the whole canvas, which is how you get artwork on transparency. content_units is the one to watch: it scales fit tolerances to the raster, which is great for parameter count and will absolutely turn a small square into a circle. timeout_sec (300) is the node's own; extra_args appends raw CLI flags last, overriding the widgets.

Outputs

svg is a STRING and a list - one string per image in the batch, so downstream nodes run once per image. Wire it straight into Inkvec Save SVG.

preview is the traced SVG rendered back to pixels and mask is that render's transparency (1 = transparent), so preview + mask through Join Image with Alpha gives you RGBA. Both need resvg-py or cairosvg; with no renderer installed the node passes the input image through and logs a note instead of failing.

Install

ComfyUI Manager: search "Inkvec" (publisher logolabs-org) and install, then restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/logolabs/inkvec-comfyui.git
pip install -r inkvec-comfyui/requirements.txt

On the Windows portable build, use python_embeded\python.exe -m pip install -r inkvec-comfyui\requirements.txt. The requirements are resvg-py, onnxruntime, huggingface_hub and einops; torch, numpy and Pillow already ship with ComfyUI.

The first trace downloads the inkvec binary for your platform from the project's GitHub releases (about 2 MB), verifies it against the release SHA256SUMS, health-checks it with inkvec --version, and caches it in ComfyUI/models/inkvec/<version>/. It is not updated automatically - delete that folder to pull a newer release. Offline, grab the archive yourself and point INKVEC_BIN at the executable.

Where people get burned

A big image times out and looks like a hang. Trace cost grows with pixel count. The node errors after timeout_sec, and the fix menu is in that message: raise timeout_sec, lower max_dim, or set time_budget - a budget expiry still gives a correct, slightly rougher trace rather than nothing.

A proxy or antivirus eats the binary. Downloading and running an executable on first use is the one part of this pack that security software objects to. If it fails, the error tells you to fetch the file by hand and point INKVEC_BIN at it.

The widgets and the binary disagree. native_alpha and content_units need inkvec 0.1.4; flip one on an older cached release and the CLI reports an unknown option. Re-download the binary (INKVEC_VERSION=0.1.4 pins it if you want it locked).

Text is traced as outlines, never as <text> - fine for a logo, surprising if you wanted selectable type. And if the source is a photograph, stop: vectorisation wants flat art, and no amount of dial-turning changes that.