FOURROOM / H-JEPA
← Original walkthroughModel explorer
Training onlySharedInference only
INPUT FROM
TRAINING LOSS · SAME CLIP, TWO PATHS
Where the target comes from
① Encode the recorded clipEpisode 0 · choose a target
Average over 192 features, all 3 next-frame pairs, and the batch.
Each level MSE + 0.18 × SIGReg(encoded states)
L1 loss + L2 loss → update model weights
Recorded images; illustrative crop. No prediction or loss value replayed.
L1 loss + L2 loss → update model weights
Click to inspect in place · hover an underlined input · drag to pan · scroll to zoom