Tencent ARC Open-Sources WorldCrafter: Implicit 3D Memory for Consistent World Exploration

On Sept 22, Tencent ARC and collaborating teams (including CUHK researchers) released WorldCrafter, a video world model with paper and code open-sourced (GitHub: TencentARC/WorldCrafter). It targets the problem that video world models struggle to respect historical observations during long-horizon interactive exploration, proposing an implicit 3D-aware memory queryable by camera poses to keep scenes consistent across minutes-long exploration and viewpoint roaming.

Confirmed

• The work comes from TencentARC and collaborators; author yshan2u posted multiple threads introducing it, CUHK researcher wbhucuk also helped spread it; paper and code are open-sourced.
• Core idea: skip explicit 3D reconstruction entirely; the target viewpoint determines how multi-view evidence is compressed into the generator's limited memory, with 3D spatial information implicitly encoded in a memory module that completes any queried view.
• The memory substrate is neither rigid 3D geometry nor plain latent memory, but a "soft skeleton" of 3D-aware latents.
• Capabilities: generating video from a single image or text prompt with historical observations; after exploring and moving objects in dynamic scenes, scene state stays persistently consistent.
• Demonstrations turned famous paintings into walkable, revisit-able 3D worlds: stepping through the door in Las Meninas, passing the arch in The Death of Socrates, viewing through the mirror in a Morisot painting; the authors use a "leave-and-return" loop test to verify persistent memory, unlike most generative worlds that "forget once you look away."

Why it matters

• Scene consistency over long horizons is a recognized challenge in video generation; achieving persistent memory without explicit rigid 3D representation offers a new paradigm for interactive world models.
• Open-source code enables reproduction and follow-up research; cross-view consistency directly benefits gaming, simulation, and embodied AI applications.

2026-09-22 ~ 2026-09-23 · 9 related posts

• [source] Tencent ARC's WorldCrafter adds implicit 3D-aware memory to video world models — TencentARC · 2026-09-22
• [source] WorldCrafter: A World Model With a 'Soft Skeleton' of 3D-Aware Latents for Persistent Memory — yshan2u · 2026-09-22
• [source] WorldCrafter: consistent video world model with implicit 3D-aware memory — yshan2u · 2026-09-22
• WorldCrafter: a video world model with memory you can walk into and revisit — CurieuxExplorer · 2026-09-22
• Tencent ARC open-sources WorldCrafter, a video world model with implicit 3D-aware memory — yshan2u · 2026-09-23

4 near-duplicate retellings: yshan2u · _akhaliq · yshan2u · yshan2u