🫥 ReMind — video generation with dynamic memory

ReMind is a Wan2.2-TI2V-5B world model taught to remember what it can no longer see. Schedule a disturbance mid-clip — an object covers the lens, the lights go out, the camera turns away — and the scene keeps evolving out of sight, so when the view comes back the state has moved on instead of resetting.

81 frames · 832×480 · 16 fps · four-step DMD rollout over seven 3-frame chunks.

First frame

What happens in the scene

Out-of-sight event

Scheduled over chunks 3–5 of 7, with two recovery chunks after it.

Generated video

Chunk prompt schedule

What the model was actually told, chunk by chunk.

Examples from the ReMind project page

Image-to-video (two occluder-recovery cases, three clean)

Camera-controlled (paired trajectories, PM-RoPE)

The occluder and the lights-out event are never painted into the pixels — they are described to the generator through chunk-local text, exactly as in the paper. The camera mode instead feeds a real trajectory (extrinsics + intrinsics) into ReMind's camera-phase RoPE; leave the trajectory file empty to use a synthetic pan-away-and-return.

📄 Paper · 💻 Code · 🤗 Weights (CC-BY-NC-4.0, research use) · base model Wan2.2-TI2V-5B