LogoLabs
LogoLabs designs custom logos and brand identities for startups, SMEs and founders, working from data. Our process is research-backed: we study how brands are drawn, what their briefs ask for and which choices hold up, and a short conversational brief becomes a set of original concepts, refined with you. The result is a logo and full brand guidelines in days, not weeks.
The research behind that work is open. Here you'll find the models, tools and datasets we build it on: generating marks, turning them into clean vector files, and understanding what makes a design brief. The tools run on your own machine or in your browser, and nothing you draw or type is sent anywhere.
Want a finished identity rather than a tool? We design custom logos with 20-page brand guidelines and full commercial rights, delivered in 24–48 hours, from €20. Start at logolabs.org or contact us.
Inkvec: raster to SVG, exactly
Inkvec is an open-source (Apache-2.0) vectoriser that turns a PNG, JPEG or WebP logo into a clean SVG. It fits every shape by minimum description length: edges placed to a fraction of a pixel, the fewest curves that match the image, gradients kept as gradients, and transparency traced as it is. In our published comparison it has the lowest colour error on every case against VTracer and Trazor, and a Fast mode trades a little of that precision for speed.
- Try it in the browser: Inkvec Studio Lite, the full editor as WebAssembly, with before-and-after comparisons on real logos.
- Desktop app and command line for Windows, macOS and Linux: releases.
- Bindings for Python, JavaScript, Go, Swift, Java, .NET, PHP and C, all generated from one options schema.
Two small models ship with it, and are released on their own too:
| Model | What it does | Size | Licence |
|---|---|---|---|
| inkvec-denoiser-001 | Removes JPEG, WebP and AI-decoder damage from flat artwork before tracing | 19.7M | Apache-2.0 |
| inkvec-sr-001 | 4x super-resolution for logos and icons (MambaIRv2-Small, fine-tuned) | 9.77M | Apache-2.0 |
Agate: a compact text-to-image model
Agate Preview 001 is a 260M-parameter text-to-image model trained from scratch in 145 GPU-hours. It scores 0.550 on GenEval with the official scorer, level with SDXL's published 0.55, and runs in under two seconds on a consumer GPU (MIT).
- Try it in the browser: Agate WebGPU, where the whole model runs on your own GPU; prompts never leave your machine.
- ComfyUI: the agate-comfyui node pack.
LogoBrief-10K: a dataset of logos and their briefs
LogoBrief-10K is 10,000 brand logos, each with its original SVG and a clean raster render, plus design annotations: open-vocabulary style tags, a one-sentence motif description, a full generated design brief, and text-region bounding boxes. Domains are sampled by web-popularity rank rather than picked for recognisable brands, and every one was checked for AI-training opt-out signals before publication. Access is gated; the card's preview images are in LogoBrief-10K-assets.
Compute
Our models were trained on Arrhenius at NAISS, through EuroHPC JU.