STORM vs Real Pong — World Model Prediction

STORMNeurIPS 2023·64x64 @ 15fps
PaperGitHub
Compare with:DIAMONDTWISTERSTORMΔ-IRISNEWIRIS
Ground Truth
Atari Pong real gameplay (ground truth frames, 64x64 @ 15fps)
Prediction
Atari Pong predicted by STORM world model (64x64 @ 15fps)
SSIMStructural Similarity Index. Measures pixel-level structural similarity. Range 0-1, higher = more similar.Higher is better
0.945
FVDFrechet Video Distance. Measures distribution-level similarity of video features. Lower = closer to real.Lower is better
42.3
LPIPSLearned Perceptual Image Patch Similarity. Measures perceptual similarity via deep features. Range 0-1, lower = more similar.Lower is better
0.0186
FramesPrediction horizon
100