Model random_depth_n1-8_long, trained with the loop depth drawn from 1–8 at every step. Each rollout starts from one real frame and chains 4 windows of 33 frames (8.8 s at 15 fps), every window conditioned on the model's own last frame. Every prompt is rolled out with 5 noise seeds. Seed 0 uses the same noise as the earlier single-seed Long Rollout pages, and every checkpoint and depth shares the same 5 seeds.
Black rows: that seed has not been generated yet. Underlined columns are depths inside the training range (N 1–8). The base and Phase-3 columns are the same for every checkpoint.
Means over the 8 cases. Speed is the time for one 4-window rollout on one H100. LPIPS vs GT is for the real prompt only and covers frames inside the episode. Seed spread is the mean pairwise LPIPS between the 5 seeds, over all prompts: higher means the seeds differ more. Motion is the mean absolute frame-to-frame change (0–255). Window columns are w1…w4. Steps after 3000 come from a rebuilt runner that regenerates the step-3000 N 1 clips bit-identically and reproduces all its numbers, except seed spread in w1 (0.180 vs the published 0.187).