Gantry

Evaluation report, WildWorld part 1

The same state-conditioned video model fine-tuned twice on WildWorld part 1, once on all 3,434 episodes as released and once with 368 episodes removed, the 349 whose picture barely changes from one frame to the next and nineteen more that run under two minutes. Same eight WildBench axes, 200 test samples on each. The edited run scored 5.2 points higher.

Fine-tuned on full
30.6%
WildBench pass rate with every one of the 3,434 episodes kept.
CI 28.4 to 32.9 · n=1,600
Fine-tuned on edited
35.9%
Same eight axes, same 200 samples, with the 349 lowest-motion episodes and nineteen short ones dropped.
CI 33.6 to 38.3 · n=1,600
Lift
+5.2 pp
From training on less data.
95% range +2.0 to +8.5
Removed
10.7%
368 of 3,434 episodes, 56.0 of the 574 recorded hours. Every kept episode keeps all four of its files.
56.0 of 574 hours · 9.76%
Two codecs agree
133
Of the 349 episodes under the motion threshold, 133 also land in the lowest tenth of the corpus on the lossless depth stream, 3.7 times the rate chance would give. A lossless encoder cannot be fooled by a dark or flat-looking scene, so when it agrees the picture really is not moving.
38.1% of the 349 · 10.3% base rate

Success rate by task

fine-tuned on full WildWorld part 1 fine-tuned on edited WildWorld part 1

A same-size random cut of 368 episodes scored 30.9%, inside the full arm's interval, so the gain is from which episodes were removed and not from removing episodes.

Gain per task

The axes that move are the ones about motion and state, Motion Smoothness and State Alignment first, which is where the removed footage was thinnest. Image Quality and Aesthetic Quality sit flat, because dropping quiet footage does not change how a single frame looks. The pooled lift clears zero, most single axes at 200 samples do not, and the chart says so.

What the edited version had removed

Measured on the whole of part 1, all 3,434 episodes and all 13,740 files, read from the public file tree with no sampling. Every episode carries rgb.mp4, depth.mkv, state.csv and skeleton.npz and not one of the four is missing anywhere. state.csv and skeleton.npz are per-frame records, so byte size stands in for length, and the two agree, skeleton.npz holds between 984 and 1,400 bytes per frame across all 3,434 episodes and tracks state.csv at r = 0.9933. On that footing the release runs from a 12.3 second episode to a 46.1 minute one around an 8.9 minute median. The number that decides the cut is a different one, how many bytes the RGB encoder spends on each frame, which is what a frame costs when it differs from the one before it. 349 episodes sit under 4.54 Mbps against a 7.57 Mbps median at the same 720p and 30 fps, and 133 of them also fall in the lowest tenth of the corpus on the lossless depth stream, 3.7 times what chance would give. Nineteen more go because they run under two minutes on an otherwise normal motion budget, in a release whose own construction pipeline is described as removing short clips. Length on its own is not a defect and the rule never treats it as one, 67 of the 92 episodes under three minutes stay, 13 of the 14 sessions over half an hour stay, and the removed part has a median length of 8.3 minutes against 9.0 minutes for what is kept. Every number on this page comes from the public file tree, the episode files themselves sit behind the manual access gate, so no frame of the release is reproduced here.

Removedepisodeshours
Under 4.54 Mbps, depth stream below its median25236.0
Under 4.54 Mbps, depth stream above its median9719.5
Under two minutes long, motion budget normal190.5
Removed36856.0
Kept3,066518.0
WildWorld part 1 · 3,434 episodes · 574 hours · Wan2.2-TI2V-5B, state-conditioned · 8 WildBench axes × 200 samples × 2 arms Wilson 95% · cut = RGB under 4.54 Mbps at 720p and 30 fps, or under two minutes long