Translations:Diffusion Models Are Real-Time Game Engines/64/en
We evaluate the impact of changing the number of past observations in the conditioning context by training models with (recall that our method uses ). This affects both the number of historical frames and actions. We train the models for 200,000 steps keeping the decoder frozen and evaluate on test-set trajectories from 5 levels. See the results in Table 1. As expected, we observe that generation quality improves with the length of the context. Interestingly, we observe that while the improvement is large at first (e.g., between 1 and 2 frames), we quickly approach an asymptote and further increasing the context size provides only small improvements in quality. This is somewhat surprising as even with our maximal context length, the model only has access to a little over 3 seconds of history. Notably, we observe that much of the game state is persisted for much longer periods (see Section 7). While the length of the conditioning context is an important limitation, Table 1 hints that we’d likely need to change the architecture of our model to efficiently support longer contexts, and employ better selection of the past frames to condition on, which we leave for future work.
