WitneyWW/perframe-1to1-world-model-rollouts

One frame predicts one frame

Autoregressive rollouts from vt_wan22_pf1_a9d, a visuo-tactile Diffusion-Forcing world model trained on a per-frame Wan2.2 latent tree.

Wan2.2's causal VAE compresses time 4x, so the usual latent tree packs four frames into every latent and a "predict the next window" model actually advances 0.67 s per step. This run encodes each frame as its own single-frame clip (iter_ = 1 + (t-1)//4 = 1, only the encoder's single-frame first chunk runs), so one latent is exactly one frame and the model predicts literally the next frame — a 167 ms step at 6 fps. An 8 s clip is 48 autoregressive steps; a 16 s clip is 96.

Videos are GT | rollout side by side. Only held-out episodes are shown.