statecliques/stereocrafter-subfixer-models

Hi-SAM text-stroke checkpoint — mirror for StereoCrafter Studio

This is an unmodified mirror of sam_tss_l_textseg.pth from Hi-SAM. I did not train it and claim no credit for it.

It is re-hosted because the original is published on Google Drive, which serves an HTML interstitial for large files rather than the bytes. That page saves perfectly happily as a .pth and then fails much later at model-load time, looking like a corrupt checkpoint — a confusing failure to hand someone who has just installed an application. Drive also throttles scripted downloads.

File

sam_tss_l_textseg.pth	122,756,163 bytes
SHA-256	1A7399FD5B031383A3776B4375332D23B952BE616A735B545B3ABB7EB89D063F

Verify before use:

certutil -hashfile sam_tss_l_textseg.pth SHA256

It holds only the trained parts. The SAM ViT-L backbone is merged in at load time from sam_vit_l_0b3195.pth, which Meta publishes directly and which is not mirrored here.

Credit

• Model and code: Maoyuan Ye, Jing Zhang, Juhua Liu, Chenyu Liu, Baocai Yin, Cong Liu, Bo Du, Dacheng Tao — Hi-SAM, Apache 2.0.
• Built on: Segment Anything, Meta AI, Apache 2.0.

@article{ye2024hi-sam,
  title={Hi-SAM: Marrying Segment Anything Model for Hierarchical Text Segmentation},
  author={Ye, Maoyuan and Zhang, Jing and Liu, Juhua and Liu, Chenyu and Yin, Baocai
          and Liu, Cong and Du, Bo and Tao, Dacheng},
  journal={arXiv preprint arXiv:2401.17904},
  year={2024}
}

Licence, and one caveat worth reading

The Hi-SAM code is Apache 2.0. These weights were trained on the TextSeg dataset, which carries its own research-use terms from its authors — see the Hi-SAM repository for how to obtain it. Nothing here grants any rights over that dataset, and if you intend to use this model commercially you should check those terms yourself.

If the Hi-SAM authors would prefer this mirror not exist, open a discussion on this repo and I will take it down.

What uses it

StereoCrafter Studio's Subtitle Fixer, which repairs the smeared depth DepthCrafter produces around burned-in subtitles: the glyphs are text to a reader but texture to a depth model, so they come back as a flickering blob that splatting then tears apart. Hi-SAM finds the glyph strokes precisely enough to pin them to a flat depth and composite them back.