Image Reconstruction on Memory Maze
0.097LPIPSCoME
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CoMESteps per frame=22026.05 | 0.097 | 0.892 | 23.07 | |
| FullSteps per frame=2, Architecture=Full-attention transformer2026.05 | 0.113 | 0.859 | 22.78 | |
| + STM+LTMSteps per frame=2, Components=STM and LTM2026.05 | 0.114 | 0.862 | 22.32 | |
| + SLTMSteps per frame=2, Component=Spatial Long-Term Memory2026.05 | 0.15 | 0.833 | 20.65 | |
| + STMSteps per frame=2, Component=Short-Term Memory2026.05 | 0.156 | 0.82 | 21.29 | |
| SSMSteps per frame=2, Architecture=Mamba blocks2026.05 | 0.158 | 0.828 | 20.62 | |
| + LTMSteps per frame=2, Component=Long-Term Memory2026.05 | 0.171 | 0.805 | 19.98 | |
| SlidingSteps per frame=2, Architecture=Sliding-window attention2026.05 | 0.183 | 0.753 | 19.02 | |
| BaseSteps per frame=2, Architecture=Diffusion Transformer (DiT)2026.05 | 0.209 | 0.771 | 19.16 |