WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

WorldCrafter stores a camera-queryable implicit 3D-aware memory that keeps video generation consistent with earlier observations across long camera moves, enabling minute-scale streaming exploration from a single image or text prompt.

WorldCrafter is a video world model built around a camera-queryable implicit 3D-aware memory, so generated scenes continue to respect prior observations over long horizons and changing viewpoints. A jointly trained memory encoder and pose-conditioned readout compress multi-view history into a fixed token set of view-specific tokens before denoising, without any explicit depth-based matching or correspondences — the requested camera pose shapes how multi-view evidence is compressed into the generator's token budget. The memory is combined with recent temporal context and few-step distillation, allowing streaming scene exploration from a single input image or a text prompt. The Hugging Face listing (2026-09-20) reports experiments on both static and dynamic scenes, with better long-horizon consistency and camera control while preserving visual quality over minute-scale rollouts; the arXiv listing (2026-09-21) describes substantial gains in long-horizon consistency and camera-control accuracy during minute-scale exploration. The two accounts agree on all substantive details, with the arXiv summary the fuller of the two.

• Core mechanism: a camera-queryable implicit 3D-aware memory conditions the video generator on history, so generation respects earlier observations across long horizons and viewpoints.
• A memory encoder plus pose-conditioned readout compress multi-view history into a fixed token set of view-specific tokens before denoising, with no explicit depth-based correspondences required.
• The memory is combined with recent temporal context and few-step distillation.
• Streaming exploration can start from a single input image or a text prompt.
• Hugging Face abstract (2026-09-20): experiments on static and dynamic scenes show better long-horizon consistency and camera control while keeping visual quality over minute-scale rollouts.
• arXiv abstract (2026-09-21): substantial gains in long-horizon consistency and camera-control accuracy during minute-scale exploration.
• The two sources agree on the method and claims; no contradictions, CVEs, or version numbers are cited in either report.