Render-test

The held-out evaluation split for low-poly render → real video: 200 paired clips, 5 s each, 21 categories.

These are the 200 clips that were held out of training — not a fresh random draw from the full corpus. The source has 9773 clips split 9573 train / 200 test, and these are that test split verbatim, so no clip here was seen during training. Re-sampling randomly from all 9773 would have mixed training data into the test set and made it useless for measuring generalization.

Which video is the input

The naming is inherited from the source corpus and is easy to get backwards:

path	role
reference_video/<category>/<uuid>.mp4	low-poly geometry render — the control signal, the input
video/<category>/<uuid>.mp4	the real video — the generation target / ground truth

render_lag_frames is the offset between the two: the render leads the real footage, so frame render_lag_frames of the render lines up with frame 0 of the real video. Taking the appearance reference as real[0] and starting the render at render[lag] makes both describe the same instant; starting the render at frame 0 instead would need a real frame from before the clip began, and the first few generated blocks come out visibly worse.

Layout

metadata.jsonl                       200 rows, one per clip
test_uuids.txt                       the 200 uuids
reference_video/<category>/<uuid>.mp4  render (input)
video/<category>/<uuid>.mp4            real (target)

Paths in metadata.jsonl are relative to the repo root and match the source corpus exactly, so code written against the full dataset runs unchanged here.

Every row carries: uuid, category, caption, render_lag_frames, and the width / height / fps / num_frames / duration of both videos.

Facts

• clips: 200, mean duration 5.00 s
• real video: 852x480 for 195/200
• render: 1344x768 for 200/200
• render_lag_frames: min -2, median 15, max 19
• total size: 0.29 GB

Categories

category	clips
egocentric_video_POV_walking	23
talking_head	17
city_building	9
cooking	9
drone_video	9
fashion_makeup	9
home_indoor	9
movie_film	9
news	9
talk_show	9
animal_insect_wildlife	8
driving	8
food_drink	8
how_to	8
nature_outdoor	8
person_crowd	8
pets	8
product_review_unboxing	8
sports_exercise	8
travel	8
tv_drama_tv_series	8