WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory
WorldCrafter uses implicit 3D-aware memory to keep video world models consistent during long camera moves.
WorldCrafter is a video world model that stores a camera-queryable implicit 3D-aware memory so generated scenes respect earlier observations across long horizons and changing viewpoints. A jointly trained memory encoder and pose-conditioned readout compress multi-view history into a fixed token set before denoising, without explicit depth matching. Combined with recent temporal context and few-step distillation, it supports streaming exploration from a single image or text prompt. Experiments on static and dynamic scenes report better long-horizon consistency and camera control while keeping visual quality over minute-scale rollouts.
- A memory encoder and pose-conditioned readout produce view-specific tokens.
- No explicit depth correspondences are required.
- Memory is combined with recent context and few-step distillation.
- Streaming exploration starts from one image or a text prompt.
- Gains cover long-horizon consistency and camera-control accuracy.
Full article136 words · extracted from huggingface.co · click to collapse
Video world models enable interactive exploration of dynamic environments, yet struggle to respect prior observations over long horizons and across viewpoints. We present WorldCrafter, a video world model that learns a camera-queryable implicit 3D-aware memory for this purpose. The key insight is to let the requested viewpoint shape how multi-view evidence is compressed into the video generator's limited token budget. Trained jointly with the video generator, a memory encoder and pose-conditioned readout module integrate historical observations into a fixed set of target view-specific tokens before denoising, without explicit depth-based correspondences. By combining this memory with recent temporal context and few-step distillation, WorldCrafter enables streaming scene exploration from a single input image or text prompt. Experiments across static and dynamic scenes show substantial gains in long-horizon consistency and camera-control accuracy while preserving visual quality during minute-scale exploration.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.24984