Extended Data Fig. 7: Detailed persistency results by item type and starting location.
From: World and Human Action Models towards gameplay ideation
a, Starting locations 1–8, numbered from left to right and top to bottom and showing the last frame for each of the eight input sequences before they are edited. b, Detailed persistency results. Each stacked bar visualizes the percentage of generated videos that were annotated as ‘Persisted’, ‘Persisted until out of frame’, ‘Did not persist’ or ‘Unusable’. On the left, we group the bars by the input/context length and the edited element. On the right, we also group by the eight (four for Vertical Jumppad) starting locations. Note that modifications in some starting locations persist less than in others. c, Examples of the more challenging start locations. We believe that this is caused by the quality and prominence of a modification. Location 2: example of a Vertical Jumppad edit that is not as persistent as in the other locations. The lower persistency is probably because of how the edit appears in an unusual place on the map and is substantially smaller than it would be in the actual game. Location 3: example of a Powercell edit that is not as persistent as in other starting locations. The inserted Powercell tends to disappear or turn into another object/character, which may be because of its small size. Location 5: example of a character edit that is not as persistent as in the other locations. The edited-in character easily disappears, probably because it is not particularly prominent or contrasting with the background.
