Labelling is the cost. We cut it.
Two papers on training segmentation models from a small fraction of the labels, six trained models anyone can download, and two demos you can try in a browser.
Two papers. Both open.
Updated 2026PixCon. Clean positives, by construction.
A contrastive branch that learns only from labelled pixels the model already gets right, so its positive set carries no pseudo-label noise. It is one switch over a DINOv2 backbone and adds nothing at inference time. On Pascal VOC with an eighth of the labels, the three-seed mean reaches 87.90 mIoU, matching the published UniMatch V2-B figure.
arxiv.org/abs/2607.03068 Paper · Under review at IEEE TPAMICW-BASS v2. Read the teacher first.
Stronger backbones broke a rule of thumb. When a DINOv2 teacher is confident about nearly every pixel, the careful pseudo-label filtering that helped weaker teachers starts to cost accuracy. A one-pass diagnostic measures that confidence and picks the rule to match, tested on Pascal VOC, Cityscapes and ADE20K across three backbone sizes.
arxiv.org/abs/2608.12773Everything is already public.
Six trained models. Weights included.
Both methods, trained on Pascal VOC, Cityscapes and ADE20K. Download the checkpoint, load it, run it. No signup, no gate.
huggingface.co/psychofict Demo · Runs in a browserTry PixCon. On your own photo.
Upload an image and see what the model segments, on Pascal VOC, Cityscapes or ADE20K classes. Hugging Face granted the demo GPU time, so it does not queue behind a shared processor.
huggingface.co/spaces/psychofict/pixcon-demo Code · Apache 2.0The training code. Configs and all.
Both repositories carry the training loop, the configs behind every published number, and the evaluation scripts. The DINOv2 backbone underneath is Apache 2.0 as well, so the whole stack is free to use commercially.
github.com/psychofict/PixCon Demo · CW-BASS v2The second demo. Built by Hugging Face.
Hugging Face’s open-source team read the released code, built a demo on top of our published weights, and offered it back. It runs on their GPU infrastructure.
huggingface.co/spaces/hugging-apps/cwbass-v2-segmentationYou have the images. You don’t have the labels.
That is the normal case in visual inspection, farming and satellite work: an archive nobody has time to annotate, and a general-purpose model that does not hit the target on your images. We take a small labelled set plus the archive you already have and train a segmentation model for your task. Training happens on your hardware, so the images stay where they are, and the model is yours to keep.
Fixed scope, fixed price, agreed before we start. Write to research@ebenworks.co with what you are trying to detect and roughly how many images you have.
What we are measuring now.
The papers are benchmarked on street scenes and everyday objects. Factory photographs are a different problem, so we ran the same methods against industrial defect images, with the splits controlled for near-duplicate photographs of the same part. Skip that control and a model scores well by recognising pictures it has already seen. The first result is published, including the part that did not go our way: the cheapest configuration matched the most elaborate one.