Interactive world models continually generate video by responding to a user's actions, enabling open-ended generation capabilities. However, existing models typically lack a 3D representation of the environment, meaning 3D consistency must be implicitly learned from data, and spatial memory is restricted to limited temporal context windows. This results in an unrealistic user experience and presents significant obstacles to downstream tasks such as training agents. To address this, we present PERSIST, a new paradigm of world model which simulates the evolution of a latent 3D scene: environment, camera, and renderer. This allows us to synthesise new frames with persistent spatial memory and consistent geometry. Both quantitative metrics and a qualitative user study show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds. We further demonstrate novel capabilities, including synthesising diverse 3D environments from a single image, as well as enabling fine-grained, geometry-aware control over generated experiences by supporting environment editing and specification directly in 3D space. Project page: https://francelico.github.io/persist.github.io
Figures & tables
Figure 1: Initialized with a single pixel frame, PERSIST evolves in an autoregressive loop in response to user actions . We first denoise the 3D environment centred on the agent in the form of a latent world-frame . Next, camera parameters are predicted with a feed-forward transformer. We then project the world to the camera plane to form a depth-ordered stack of world latents . Finally, pixel latents are denoised, using pixel-aligned 3D information from the world latents stack as guidance.
Figure 2: PERSIST enables long-horizon spatial memory by modelling the dynamics of a 3D world-frame around the agent. Camera parameters then act as memory look-up key, fetching relevant features from the world frame via a geometric projection (here visualized as the coloured voxels).
Figure 3: PERSIST can be initialized with a single RGB frame ( , row 1), or with a single RGB and world frame ( + , row 2). We visualize the world-frames and videos produced by an autoregressive rollout of 600 timesteps. Even with a single RGB frame for initialization, PERSIST can generate cohesive and evolving worlds.
Figure 4: World frame w features are projected to screen-space to obtain the depth-ordered stack of features w~2D and linear depth information d .
Figure 5: Video frames generated over 600 timestep episodes by PERSIST (Ours), Oasis ( Decart et al., 2024 ) and WorldMem ( Xiao et al., 2025a ) .
Figure 6: FID scores compared to ground truth over 600 frame episodes. PERSIST configurations remains stable, while baselines relying on pixel-histories degrade rapidly.
Method
200 Frames
400 Frames
600 Frames
Oasis
409
687
875
WorldMem
358
–
–
No-3D-Upscale
216
231
247
Camera-GT
161
152
152
PERSIST-S
159
170
179
PERSIST
129
141
148
Table 1: FVD ( ↓ ) evaluated across different episode lengths (200, 400, and 600 frames). We compare our base configuration (PERSIST) against pixel-history baselines and various model ablations. PERSIST +w0 is discussed in Section 6.2 .
Method
VF ↑
3D ↑
Temp ↑
Overall ↑
Oasis
2.1 ± 0.1
1.9 ± 0.1
1.8 ± 0.1
1.9 ± 0.1
WorldMem
1.7 ± 0.09
1.7 ± 0.09
1.5 ± 0.08
1.5 ± 0.07
PERSIST-S
2.8 ± 0.1
2.7 ± 0.1
2.5 ± 0.1
2.6 ± 0.09
PERSIST
2.8 ± 0.09
2.5 ± 0.09
2.5 ± 0.09
2.6 ± 0.08
PERSIST +w0
3.2 ± 0.1
2.8 ± 0.1
2.8 ± 0.1
3.0 ± 0.1
Table 2: Mean user ratings (score range: 1-5) for videos generated from different PERSIST configurations and baselines. Bold numbers indicate the best performing methods. VF = Visual Fidelity, 3D = 3D Consistency, Temp = Temporal Consistency, Overall = Overall Score.
Figure 7: PERSIST’s explicit 3D representation allows us to edit the world-state during an episode, enabling intuitive fine grained control over generations. We demonstrate this above with both global edits to terrain and biome, as well as the placement of smaller assets such as trees.
Figure 8: PERSIST’s 3D state enables collision modelling with out-of-view elements (top: player moves backwards into a tree at t=50 ). Since the 3D state is dynamic, it evolves even when unobserved (middle: a cave filling with water). This allows off-screen events to produce emergent on-screen effects (bottom: water flowing onto the player at t=150 ).
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Table 11
Figure 9: User Study Interface. A screenshot of our web-based platform where participants evaluate video pairs based on temporal and spatial consistency.
Figure 10: Score differences during head-to-head comparisons across evaluation metrics.
Method
Per Frame Visual Fidelity ↑
3D Consistency ↑
Temporal Consistency ↑
Overall Score ↑
Oasis
2 ± 0.2
1.6 ± 0.2
1.5 ± 0.2
1.6 ± 0.2
WorldMem
1.7 ± 0.2
1.5 ± 0.1
1.3 ± 0.1
1.4 ± 0.1
PERSIST-S
2.8 ± 0.2
2.6 ± 0.2
2.3 ± 0.2
2.5 ± 0.2
PERSIST
3.0 ± 0.2
2.5 ± 0.1
2.2 ± 0.1
2.5 ± 0.1
PERSIST +w0
3.4 ± 0.2
3.0 ± 0.2
2.7 ± 0.3
3.1 ± 0.2
Appendix
Table 10: Per set user study scores: circle around .
Method
Per Frame Visual Fidelity ↑
3D Consistency ↑
Temporal Consistency ↑
Overall Score ↑
Oasis
2.3 ± 0.2
2.2 ± 0.2
2.0 ± 0.2
2.0 ± 0.2
WorldMem
1.5 ± 0.2
1.5 ± 0.2
1.2 ± 0.1
1.3 ± 0.1
PERSIST-S
3.0 ± 0.2
2.9 ± 0.2
2.8 ± 0.2
3.0 ± 0.2
PERSIST
2.9 ± 0.2
2.8 ± 0.2
2.8 ± 0.2
2.8 ± 0.2
PERSIST +w0
3.6 ± 0.2
3.1 ± 0.2
3.3 ± 0.2
3.4 ± 0.2
Appendix
Table 11: Per set user study scores: free play .
Method
Per Frame Visual Fidelity ↑
3D Consistency ↑
Temporal Consistency ↑
Overall Score ↑
Oasis
1.9 ± 0.2
2.0 ± 0.2
2.0 ± 0.2
1.9 ± 0.2
WorldMem
1.7 ± 0.2
1.8 ± 0.2
1.6 ± 0.2
1.6 ± 0.1
PERSIST-S
3.0 ± 0.2
2.7 ± 0.2
2.7 ± 0.2
2.6 ± 0.2
PERSIST
2.7 ± 0.2
2.4 ± 0.2
2.4 ± 0.2
2.5 ± 0.1
PERSIST +w0
3.0 ± 0.2
2.6 ± 0.2
2.6 ± 0.2
2.8 ± 0.2
Appendix
Table 12: Per set user study scores: backwards and look around .
Method
Per Frame Visual Fidelity ↑
3D Consistency ↑
Temporal Consistency ↑
Overall Score ↑
Oasis
2.4 ± 0.2
2.0 ± 0.2
1.8 ± 0.2
1.9 ± 0.2
WorldMem
2.1 ± 0.2
2.2 ± 0.2
1.9 ± 0.2
1.9 ± 0.2
PERSIST-S
2.2 ± 0.2
2.3 ± 0.2
2.1 ± 0.2
2.3 ± 0.2
PERSIST
2.6 ± 0.2
2.3 ± 0.2
2.3 ± 0.2
2.4 ± 0.1
PERSIST +w0
2.7 ± 0.2
2.4 ± 0.2
2.3 ± 0.2
2.4 ± 0.2
Appendix
Table 13: Per set user study scores: forward and look around .
Method
Time per frame (ms) ↓
PERSIST-S
2672.4
Oasis
3007.4
WorldMem
6387.1
Appendix
Table 14: Inference speed comparison against baselines. All tests were conducted on a single A100 GPU with a batch size of 1 and 20 denoising steps.
Component
Time per frame (ms)
World frame
1538.4
Pixel frame
1111.0
Camera state
21.9
Projection
0.63
Appendix
Table 15: Component-wise breakdown of generation time per frame for PERSIST-S (20 denoising steps).
Figure 11: Conditioning on the ground truth camera trajectory results in physical inconsistency with the world-frame, causing the agent to phase through a wall.
Figure 12: PERSIST generates diverse but coherent initial world frames from a single RGB conditioning observation. Each row corresponds to a specific input RGB frame, while each column depicts the initial world frame sampled from Wθ during a specific generation episode.
Figure 13: Generation artifacts and subsequent recoveries over a 2000-step (83 seconds) episode generated with PERSIST. The 3D representation drifts locally, causing individual blocks to appear and disappear from the agent’s view. However we find that the 3D representation remains globally coherent and has a net stabilising effect on generation, allowing Pθ to recover from visual artifacts. This lets PERSIST generate rollouts lasting thousands of steps, even if glitches become more frequent over time due to autoregressive drift.
Recent advances in video world models have enabled plausible interactive environments, yet long-horizon generation places a fundamental demand on visual memory: the previously observed world must remain consistent as the model continues to generate. Existing autoregressive approaches often implicitly rely on the generative model itself to retain such information, leading to spatial memory decay beyond the active context. We introduce World-T3, a camera-controlled video world model with a test-time training memory that consolidates past observations into dedicated model parameters. To robustly update this memory, we propose a recursive least-squares rule that incrementally absorbs novel visual features via online pose-keyed memory updates. Our TTT memory reads out previously observed scene features from the fast-weight state by camera pose, enabling consistent generation with constant storage and per-step computation regardless of rollout length. To facilitate real-time interactive generation with this robust memory, we further distill the autoregressive model into a few-step generator via dual-teacher distribution matching. On a long-horizon revisit benchmark, World-T3 raises revisit PSNR from 13.52 to 18.31 dB and reduces absolute rotation error from 3.344 to 1.075 relative to the strongest baseline on both metrics, while sustaining 20.6 FPS at 1280 x 704 on a single NVIDIA B200 GPU.
Junyi Wu, Zekun Li, Chenhongyi Yang +9
Meta Reality Labs · University of Illinois Chicago · Brown University
We present WorldDirector, a highly controllable video world model framework designed for persistent dynamic object memory and unrestricted viewpoint exploration. Unlike existing world models that entangle physical dynamics with pixel rendering and rely on continuous visual observation to sustain motion, our framework explicitly decouples semantic motion orchestration from visual generation. By leveraging an LLM to coordinate 3D trajectories with camera movements and subsequently employing these orchestrated trajectories as control signals for video generation, our approach ensures strict physical logic and appearance stability, successfully preserving the exact visual identities of dynamic entities even when they re-enter the scene after prolonged periods out of view. Experimental results demonstrate that our method supports the synthesis of complex and extended events with unprecedented controllability and persistent dynamic object memory. Project Page: https://worlddirector.github.io/
Recent advances in video generation enable a new paradigm for 3D scene creation: generating camera-controlled videos that simulate scene walkthroughs, then lifting them to 3D via feed-forward reconstruction techniques. This generative reconstruction approach combines the visual fidelity and creative capacity of video models with 3D outputs ready for real-time rendering and simulation. Scaling to large, complex environments requires 3D-consistent video generation over long camera trajectories with large viewpoint changes and location revisits, a setting where current video models degrade quickly. Existing methods for long-horizon generation are fundamentally limited by two forms of degradation: spatial forgetting and temporal drifting. As exploration proceeds, previously observed regions fall outside the model's temporal context, forcing the model to hallucinate structures when revisited. Meanwhile, autoregressive generation accumulates small synthesis errors over time, gradually distorting scene appearance and geometry. We present Lyra 2.0, a framework for generating persistent, explorable 3D worlds at scale. To address spatial forgetting, we maintain per-frame 3D geometry and use it solely for information routing -- retrieving relevant past frames and establishing dense correspondences with the target viewpoints -- while relying on the generative prior for appearance synthesis. To address temporal drifting, we train with self-augmented histories that expose the model to its own degraded outputs, teaching it to correct drift rather than propagate it. Together, these enable substantially longer and 3D-consistent video trajectories, which we leverage to fine-tune feed-forward reconstruction models that reliably recover high-quality 3D scenes.