The most embarrassing moment for a video world model is not a visual glitch: it is when you steer the camera in a circle, come back, and the room you generated minutes ago has turned into a different one — the model simply forgot. On September 21, TencentARC published WorldCrafter on arXiv. The paper climbed to #2 on Hugging Face Daily Papers the next day (111 upvotes), the GitHub repository has gathered 218 stars, and both code and weights are open-sourced.
Amnesia Is the Real Bottleneck
Open-source world models have been arriving in waves this year: JD's EchoWM pushes omni-modality, ForgeWM pushes one-step denoising at 72FPS, SolarWM trains on 1.43 million videos. Most of them, however, steer around the same problem: long-horizon revisit consistency. Video generation leans on the most recent seconds of context; once historical observations slide out of the context window, scene information is gone.
WorldCrafter's answer is a camera-queryable implicit 3D-aware memory attached to the generator: a memory encoder trained jointly with the video generator, plus a pose-conditioned readout module, compresses past multi-view observations into a fixed set of target-view-specific memory tokens injected before denoising. The "implicit" part matters — no explicit depth estimation, no geometric warping; the requested viewpoint decides which historical evidence gets compressed into the limited token budget.
The Official Scorecard
The project page reports a benchmark spanning 145 scenes and 725 camera trajectories, where WorldCrafter improves long-horizon revisit consistency and camera-control accuracy over all evaluated baselines while achieving the highest overall VBench score. Note that these numbers come from the team's own report; independent third-party replication has not appeared yet.
On the engineering side there are two models: WorldCrafter-Base and the few-step-distilled WorldCrafter-Fast. Fast supports image-to-video and text-to-video at 384×640 resolution, generating in 33-frame chunks; camera trajectories are described with NumPy matrices, and an action-style DSL (forward, yaw_left) is also provided. The environment requires Python 3.11, PyTorch 2.10 and CUDA 12.8; weights live on HuggingFace (TencentARC/WorldCrafter-Base and -Fast).
So What
Two things to watch. First, the README states plainly that the interactive demo is "currently being debugged"; the single-GPU demo runs Fast image-to-video with compilation enabled — smooth real-time exploration is still some distance away, and whether distillation buys enough speed for streaming interaction awaits community testing. Second, the axis of competition in world models is shifting from generation quality to memory architecture: whoever can remember the world it generated deserves to be called a world model.
Reference: https://arxiv.org/abs/2609.24984 · github.com/TencentARC/WorldCrafter · huggingface.co/papers/2609.24984