The problem
Good cutouts cost a subscription — and your images
The usual tools watermark the free tier and upload your originals to somebody else's cloud. This one runs the models locally and charges nobody.
kaicorplabs.com / projects / pixelforge
project open source · self-hosted
Cut the background. Keep the pixels. Keep nothing else.
Background removal and image vectorisation running on our own machine: full resolution out, no watermark, and your image is gone the moment the job ends.
Who it is for
Design work, shops, one-off jobs
Product photos that need clean cutouts; logos that need to scale to a billboard. Four segmentation models to choose from, plus vectorisation with real controls.
Your data
Alive for one request, then gone
Uploads live in a private temporary space for exactly the duration of the job and are deleted in a finally block — with sweeps for anything orphaned. Nothing is stored, so there is nothing to leak.
Honest limits
Real CPU, real queues
Neural networks on CPU cost seconds per job: two run at once, the rest queue, uploads cap at 50 MB, and every job requires an account — an open endpoint would be free compute for strangers.
Run your own copy
First start downloads the four ONNX models (~400 MB) with their sha256 pinned — if upstream ever served different bytes, it refuses to boot. The heaviest image of the family.
docker pull ghcr.io/ulzuhan/pixelforge:latest # or: clone the repo and `docker compose up -d --build` # DEPLOYMENT.md in the repository covers the rest.
The image is built from a tagged commit by CI, scanned for known vulnerabilities before release, and published with provenance and an SBOM. Or use ours: open pixel.kaicorplabs.com and ask for an account — a person approves it.