Each model makes 4 windows of 33 frames in a row. Window 1 starts from the real last frame and is the gallery clip. Each later window starts from the last frame the model itself generated, with the same prompt and fresh noise. 133 frames at 15 fps. The looped columns belong to the checkpoint selected below; its training step is in the column header.
Generation time for 4 windows on one GPU of the 2×H100 SXM box, while other jobs shared it: base 31 s, fine-tuned 31 s; looped, 4500 steps: N = 1 20 s, N = 2 19 s, N = 4 27 s, N = 8 39 s, N = 16 70 s; random depth N 1–8, step 2500: N = 1 15 s, N = 2 18 s, N = 4 25 s, N = 8 37 s, N = 16 65 s. Depths marked (trained) in the column headers are the ones the selected checkpoint was trained at; the others are extrapolation. Frames from AgiBot World Beta (CC BY-NC-SA 4.0: attribution, non-commercial, share-alike).