Backward → Yaw left · r = 2
TL;DR
ActionSplice edits an in-progress diffusion state when the requested action changes during sampling. CST-R retargets the active chunk; CST-T preserves the completed prefix and applies the new action to the remaining suffix. Both resume the frozen world model from the current solver step instead of restarting generation.
01 / Method
Transport the state. Keep the model frozen.
CST edits the model-specific clean prediction, reconstructs the exact scheduler-consistent state, and resumes ordinary denoising.
01Interrupt
Capture the old-action active state.
→02Transport
Predict the matched clean-state correction.
→03Reconstruct
Apply the native scheduler realization.
→04Resume
Execute only the remaining frozen-model NFEs.
02 / Qualitative
CST-R applies the revised action within the active chunk.
Backward → Forward · r = 2
Forward → Yaw left · r = 2
Repeated updates
CST-R handles repeated in-flight action updates.
Matched HY-WM1.5 rollouts with repeated action interruptions at r = 2.
Five sequential action updates · r = 2
Two sequential action updates · r = 2
CST-T / Qualitative
CST-T: Within-chunk action transport
Matched Wait controls versus CST-T · r = 2.
03 / Results
Evaluation highlights.
Quantitative results reported in the current manuscript draft.
CST-R · rollback fidelity
61.5%lower LPIPS on minWM
75.9%lower LPIPS on HY-WM1.5
Relative to direct condition swapping.CST-T · suffix fidelity
56.1%lower LPIPS on minWM
77.5%lower LPIPS on HY-WM1.5
Relative to direct condition swapping.CST-T · pixel-ready speedup
2.73×minWM
1.69×HY-WM1.5
Relative to waiting.HY-WorldPlay · CST-R
25.66PSNR
0.6902SSIM
0.1337LPIPS
Against the original rollout.04 / Citation
Citation
Pending
BibTeX will be added with the paper release.