Datasets:

---

nikopueringer

/

CorridorKeyDataset

Like 18

This restructures the dataset so every clip uses a consistent format. No EXR content changes (it's a pure rename) so HF won't increase storage costs since the new commits point to the same blobs. The only new uploads are the preview jpgs and the manifest (about 160 MB).

What changed

• Every clip is now Clip_0000/ to Clip_1078/, each with Alpha/, BG/, FG/, Input/ holding Frame_0000.exr onward, numbered from 0. The old folders used five different naming schemes.
• Six folders accidentally held two different clips (Lucas2 (56), Lucas2 (65), Lucas2 (134), Shane (80), Shane (81), Shane (82)). They are now split into two clips each.
• Removed a few broken frames, the stray .lnk, upload.log, and upload_to_hf.py.
• Added preview.jpg to every clip: 10 evenly spaced Input frames over their Alpha frames.
• Added manifest.json with one entry per clip: name, author, frames, md5, description, and source. (source is the original folder name so if you had something hardcoded to the old directory names you can use the manifest to remap it)
• Updated the README to document everything

Some clips are duplicates. I did not want to delete anything so I left them in. The md5 hash can be used to deduplicate exact copies but some clips seem to have been reencoded so their hashes don't match despite being duplicates.

Every local file was hashed against the LFS pointers before pushing to confirm the renames point at the right objects.