MaLissa / README.md
2mm's picture
Add MaLissa GT/LQ paired dataset (16 clips, tar-sharded)
c458dac verified
|
Raw History Blame Contribute Delete
969 Bytes
metadata
license: cc-by-nc-4.0
task_categories:
  - image-to-image

MaLissa

Paired low-quality / ground-truth video-frame dataset for frame restoration.

Contains 2,060 curated pairs (1,619 for training and 441 for validation), selected from 16 raw video clips (3,988 frames total). Each pair is a degraded input frame (lq) and its aligned ground-truth (gt).

Structure

Data is sharded as one tar per clip:

train/videoN.tar
valid/videoN.tar

Each videoN.tar extracts to:

videoN/
  gt/00000.png ...   # ground truth
  lq/00000.png ...   # low-quality input
  • train clips (13): video1, video3, video5, video6, video7, video8, video9, video10, video11, video12, video13, video14, video16
  • valid clips (3): video2, video4, video15

Usage

huggingface-cli download 2mm/MaLissa --repo-type dataset --local-dir MaLissa
cd MaLissa
for f in train/*.tar valid/*.tar; do tar -xf "$f" -C "$(dirname "$f")"; done