Causal autoregressive video generation

Teaching Video Generators to Remember: Eliciting Dynamic Memory for Out-of-Sight State Evolution

ReMind is a causal video generator that keeps world state alive while it is hidden—behind camera motion, physical occluders, or changing illumination—and brings it back coherently when the view returns.

01 · Out-of-sight state evolution

World state keeps evolving out of sight.

The visible scene can leave the frame or disappear behind an interruption. ReMind continues the underlying dynamics instead of reconstructing a frozen past.

Camera pan

Hidden content evolves while the view moves away and remains consistent when the camera reveals it again.

Occluder and light toggle

Physical events continue through temporary occlusion and changes in illumination.

02 · Degradation controllability

Control when the interruption happens.

A shared pouring scene is generated cleanly or with light and occluder events placed early, in the middle, late, or twice.

Clean
Light · early
Light · middle
Light · late
Occluder · double
Occluder · early
Occluder · middle
Occluder · late

03 · Occluder diversity

Change the occluder, not the hidden world.

The same temporal event supports varied physical occluders, from cardboard and fabric to a full light toggle.

Clean
Brown cardboard box
Cardboard sheet
Folded bed sheet
Tablecloth
Closed curtain
Velvet curtain
Shower curtain
Corrugated cardboard
Light toggle

04 · Out-of-distribution generalization

Recovery is not an occluder lookup.

In this V2V setting, we introduce held-out pure-white fog at test time that becomes fully opaque—an occluder never seen during training. ReMind still recovers the continuously evolved scene, showing that it learned persistent dynamics rather than memorizing interruption patterns.

Pouring and foam · Reference
Pouring and foam · ReMind
Granular pile · Reference
Granular pile · ReMind
Sauce spreading · Reference
Sauce spreading · ReMind
Batter spreading · Reference
Batter spreading · ReMind

05 · Clean video generation

Memory without a disturbance.

ReMind also produces clean, undisturbed video across animals, people, fluids, fire, granular material, and everyday actions.

Dog in motion
Aquarium
Waterfowl
Candle flame
Two-person motion
Running
Fish school
Waterfall
Tennis
Open flame
Cooking
Human motion
Aquarium feeding
Conversation
River flow
Dry-ice fog
Sewing
Granular pour

06 · Method

A causal model built to remember.

ReMind combines geometry-aware history, disturbance-robust training, and purpose-built dynamic video data to make out-of-sight state evolution work.

1

PM-RoPE

A 3D rotary position embedding controls camera motion while telling the current frame its temporal and camera-space distance to every history frame.

→ the key geometry for out-of-sight state evolution

2

Node Drop

A node-drop training mechanism teaches the model to ignore occlusions and other disturbances, then recover the evolving scene state when visibility returns.

→ robust state recovery through interruptions

3

Dynamic data engine

InSpatio-World provides large-scale synthetic camera motion, Helios provides strongly time-dynamic videos, and a curated occluder set provides diverse disturbances.

→ camera, dynamics, and occlusion at scale

The release includes complete 480p teacher-forcing and SF-DMD training recipes for Wan 2.2 TI2V 5B and Wan 2.1 T2V 1.3B.

Code ↗

07 · Citation

Citation

If you find ReMind useful, please cite our paper.

@article{xu2026teaching,
  title={Teaching Video Generators to Remember: Eliciting Dynamic Memory for Out-of-Sight State Evolution},
  author={Xu, Tianshuo and Xie, Yichen and Meng, Depu and Peng, Chensheng and Herau, Quentin and Jiang, Bo and Hu, Yihan and Zhan, Wei},
  journal={arXiv preprint arXiv:2605.25333},
  year={2026}
}