---
license: apache-2.0
library_name: pytorch
pipeline_tag: image-feature-extraction
---

## Checkpoints of ScaleLSD 

This repository contains the checkpoints as described in the paper [ScaleLSD: Scalable Deep Line Segment Detection Streamlined](https://huggingface.co/papers/2506.09369).
For code, see https://github.com/ant-research/scalelsd.

There are two checkpoints of our model, [scalelsd-vitbase-v1-train-sa1b.pt](https://huggingface.co/cherubicxn/scalelsd/blob/main/scalelsd-vitbase-v1-train-sa1b.pt) and [scalelsd-vitbase-v2-train-sa1b.pt](https://huggingface.co/cherubicxn/scalelsd/blob/main/scalelsd-vitbase-v2-train-sa1b.pt).\

The main differences between v1 and v2 models are summarized as follow

| Model Name    | Use of LayerScale | Optimizer |
| -------- | ------- |------- | 
| [scalelsd-vitbase-v1-train-sa1b.pt](https://huggingface.co/cherubicxn/scalelsd/blob/main/scalelsd-vitbase-v1-train-sa1b.pt)  | No    | ADAM |
| [scalelsd-vitbase-v2-train-sa1b.pt](https://huggingface.co/cherubicxn/scalelsd/blob/main/scalelsd-vitbase-v2-train-sa1b.pt)  | Yes    | ADAMW |

We recommend the users use the v2 model by default because it works much better than v1 (the model evaluated in our paper)