The ViTeX-Edit-14B results in v1 of the paper were produced with the removed inference_example.py, which ignored the glyph video; later versions report results from the GitHub code.
Citation: url points to the arXiv paper; protect {ViTeX-Bench} in the title
Point self-clone URL and download command to renamed repo
Anonymous Authors
74f6150
README's git clone now targets huggingface.co/ViTeX-Bench/ViTeX-Edit-14B
and cd's into ViTeX-Edit-14B; inference_example.py's diagnostic for a
missing base model now suggests
huggingface-cli download ViTeX-Bench/ViTeX-Edit-14B
The bundled weights filename (vitex_14b.safetensors) is unchanged
since renaming it would invalidate every existing local clone.
Rename displayed model name to ViTeX-Edit-14B in the model card
Anonymous Authors
9b9fe26
README heading, file-tree comments, and the Composite-variant
section heading all switched. Same change applied to the docstrings
of inference_example.py and make_corp_baseline.py and to the
'Loading ... trained weights' log line. Repository URL, the bundled
weights filename (vitex_14b.safetensors), and the local clone target
directory are intentionally unchanged.
Match the Dataset and Benchmark code READMEs: same five-link header
bar, anonymous-release blockquote, then Specs / Repository / Inputs /
Usage / Composite variant / Limitations / License / Citation. Drops
the redundant "self-contained" paragraph (already implied by the
repository tree) and inline training stage details (those live in the
paper); keeps every command needed to run inference end-to-end.
Unify companion-repo links in the model card header
Anonymous Authors
8fc9a54
Replace the single Project-page line with the same four-link bar used
on the project page and leaderboard:
🌐 Project page · 📊 Dataset · 🧪 Benchmark code · 🏆 Leaderboard
The model card itself fills the "Model & Inference code" slot, so it
is not linked back to itself.
Rename '(Corp)' variant to '(Composite)' to match the paper
Anonymous Authors
3ceb528
The locality-preserving post-processing variant is referred to as
"ViTeX-14B (Composite)" in the paper and on the leaderboard. Update the
model card and the make_corp_baseline.py docstring to match. The
script filename and the default output directory (ViTeX-14B_Corp/) are
left as legacy identifiers so existing eval pipelines keep working.
`make_corp_baseline.py` composes a raw ViTeX-14B prediction's text region back
onto the source video with two per-frame operations:
1. Reinhard mean-variance LAB color matching on a 20-px band just outside
the mask, so the predicted glyphs match the source's local lighting.
2. Signed-distance feathered alpha compositing (4-px feather centered on
the mask boundary), so the boundary has no visible seam.
Inside the mask the result is the predicted glyphs (color-matched); outside
the feather the result is byte-identical to the source. SeqAcc / CharAcc /
TTS are within ~0.01 of raw ViTeX-14B (the predicted text region is
unchanged), while PSNR / SSIM / LPIPS / DreamSim jump to near-Identity.
CPU-only, ~5 min on 8 workers for the 157-clip evaluation split. Reference:
appendix G of the ViTeX-Bench paper.
