LibreLingBotVisionl-sem

LibreLingBotVisionl-sem

LingBot-Vision ViT-L/16 (300M backbone) self-supervised backbone with a LibreYOLO-trained 1x1 dense head for ADE20K 150-class semantic segmentation at 512x512.

Source

The backbone derives from robbyant/lingbot-vision ("Vision Pretraining for Dense Spatial Perception", Fu et al., 2026, arXiv:2607.05247), weights from robbyant/lingbot-vision-vit-large. Copyright (c) 2026 Robbyant. Licensed under the Apache License 2.0.

Modifications

The backbone tensors are unchanged (the LibreYOLO port is parity-verified at max_abs_diff == 0 against the reference implementation). The dense head is a 1x1 convolution over the frozen patch-token grid, trained by LibreYOLO on ADE20K following the linear-probing protocol of the upstream technical report (ADE20K val mIoU 51.4 at 512x512, single scale; upstream linear-probe reference 52.8). Conversion and metadata wrapping: weights/convert_lingbotvision_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreLingBotVisionl-sem.pt")
results = model.predict("image.jpg")
results.save("out.jpg")

Fine-tune on your own semantic dataset (head-only by default, matching the linear-probe recipe; pass freeze_backbone=False for a full fine-tune):

model.train(data="your_semantic.yaml", epochs=20)

License

Apache License 2.0. See the LICENSE and NOTICE files in this repository.

Dataset Note

ADE20K has separate dataset terms and is not redistributed in this repository.