Recent interactive video world models generate visually compelling videos from a single image, following user-provided camera motion and text prompts. However, such low-level control demands substantial user effort to realize a high-level intent, and the generated world struggles to remain persistent as the exploration proceeds. In this work, we present AgiWorld, an agentic video world model that explores and builds a persistent world from a single image and a free-form instruction. For high-level planning, our agent AgiWorld-Agent designs the world and the journey through it in a world state and directs the exploration. For low-level realization, our video world model AgiWorld-WM generates the video chunk by chunk and stores it with its geometry in a spatial memory, from which the agent updates its world state. This closed loop keeps the world faithful to the user's intent and keeps the agent's world state consistent with the generated content throughout the exploration. We co-design the two levels around the shared spatial memory and an interface of metric camera trajectories and structured prompts. In AgiWorld-WM, to handle occlusion and accumulated geometric errors in the spatial memory over long rollouts, we propose occlusion-aware history selection, which retrieves past views by surface visibility, and reliability-based memory maintenance, which prefers reliable geometry and earlier observations. AgiWorld-WM achieves state-of-the-art overall performance on world revisiting and exploration benchmarks, and qualitative results and ablations show that AgiWorld builds persistent worlds from free-form instructions over minute-long journeys. Persistence thus arises not from the video model alone but from its closed loop with the agent.
Figures & tables
Figure 1: Persistent worlds built by AgiWorld. Given an input image and a free-form instruction, AgiWorld-Agent designs the world and the journey through it, and directs AgiWorld-WM with a camera trajectory and a text prompt for each chunk. AgiWorld-WM generates the video autoregressively and uses a point cloud as its spatial memory, which in turn is fed back to AgiWorld-Agent for further exploration. We show the rendered spatial memory with the camera route planned by the agent, and frames sampled from the long rollout videos. The results demonstrate the 3D persistence of the generated worlds as well as semantically meaningful exploration routes of the agent.
Figure 2: Overview of AgiWorld. The director D turns the world state Sk into ck and pk . AgiWorld-WM generates vk from the spatial history read from its memory (purple, two of five shown) and the temporal history (gray), and writes vk with its depth Dk into the memory, which the state update U reads to write Sk+1 .
Figure 3: Pipeline of AgiWorld-Agent. The state update U builds the initial world state S1 from I0 and g0 . For each chunk, the director D plans ck and pk , AgiWorld-WM generates vk , and the director reviews it. A chunk that fails is generated again, and a chunk that passes is used by U to write Sk+1 .
Figure 4: Data pipeline of AgiWorld-WM. (1) Data sources: four real-world and five synthetic datasets, filtered with metadata and a VLM. GT marks the datasets with ground-truth metric poses. (2) Data annotations: metric depth, camera poses and intrinsics from Depth Anything 3 ( Lin et al., 2025 ) , and a structured caption for every 5-second clip from Qwen3-VL ( Bai et al., 2025a ) . (3) Data filtering: geometric gates discard segments with unreliable pose and depth annotations. (4) Data sampling: clips are labeled with a camera motion class, and rare classes are sampled more often. The bars show the share of each class before and after rebalancing.
Figure 5: World Explore benchmark. Left: the first frames of 20 of the 40 scenes. Right: distance of the camera from its starting position along the target trajectories of DL3DV-Evaluation and World Explore. DL3DV-Evaluation trajectories move away and come back to a median distance of 1.6 m from the start, whereas World Explore trajectories keep moving away and end at a median distance of 18.1 m.
Figure 6: Camera-controlled long video generation comparison on DL3DV-Evaluation and World Explore. The first column shows the input image with a top view of the target camera trajectory.
Figure 8: Qualitative ablation of AgiWorld-Agent on two examples, each shown with its input image and abbreviated instruction. Each row shows frames sampled in order along the video of one variant.
Figure 9: Replacing the world model of AgiWorld. All rows use the same AgiWorld-Agent, input image, instruction and initial world state, and differ only in the world model: LingBot-World, Matrix-Game 3.0, Lyra 2.0 and AgiWorld-WM from top to bottom.
Figure 10: All decisions and state changes of AgiWorld-Agent for one chunk, from the temple example in Figure 1 : the director plans the chunk, AgiWorld-WM generates it, the director reviews it, and the state update turns Sk into Sk+1 .
Step
Called
Writes
Director D
Trajectory
At initialization, and whenever the route has to change (replan)
Waypoints of the next 60 s
Fix
After Trajectory, when the solved path fails a check
Corrected waypoints
Prompt
After Trajectory and Fix
Start, middle, and end of the prompt of each chunk
Review
After every chunk
Verdicts, small layout edits, stage updates, decision for the next chunk
State update U : reconstruction
Table 3: The steps of AgiWorld-Agent, grouped by the function they carry out. Each step has its own guidelines and output schema.
Field
Meaning
Layout element
Name
What the element is, with a fixed feature of the place when several are alike, e.g., “offering table on the passage side”. Words that depend on the camera, such as left or ahead , are not allowed.
Kind
Surface, terrain, structure, object, opening, or zone. Surfaces and terrain form the ground, and structures and objects are solids.
Shape
Polygon, polyline, path, box, circle, or point, with metric coordinates in the (x,z) plane and an optional elevation per vertex for ramps and stairs.
Elevation, height
Height of the base above h=0 , and vertical extent.
Appearance
One line that describes what the element looks like.
Table 4: Fields of the world state.
Tool output
Content
Read by
Depth map of I0
The depth of I0 as a color map with a metric scale.
Initial reconstruction
Point cloud and samples of I0
The leveled point cloud of I0 from above, and I0 with a grid of points labeled with their (x,z,h) .
Initial reconstruction
Layout check
For each element, whether its position and height agree with the depth and the point cloud of I0 .
Initial reconstruction
Layout projection
The layout drawn as wireframes into I0 or into the last frame of a chunk.
Initial reconstruction, Revision, Review
Floor plan
The layout over the accumulated point cloud, with the executed path and the planned path.
All steps except Prompt
Trajectory check
Collisions, eye-height and support violations, and violations of the motion bounds on the solved path.
Fix
Table 5: Outputs of the geometric tools and the steps that read them.
Figure 11: Renders prepared by the geometric tools in one iteration, redrawn from the intermediate results of the temple example in Figure 1 , with the point cloud in its image colors and all confirmed elements in one color. The iteration generates the chunk from 30 to 35 s, in which the camera walks into the temple hall, and its path comes from a replan after the previous chunk. Top, before the chunk is generated: (a) the floor plan of the layout with the remaining planned waypoints, read by the trajectory step, (b) the solved path, read by the fix step, and (c) the planned views at 30 s and 35 s, read by the prompt step. Bottom, after the chunk is generated: (d) frames of the chunk, (e) the layout projected onto its last frame, and (f) the floor plan of the accumulated point cloud with the executed path, all read by the review. Confirmed elements are drawn solid and imagined elements dashed.
Behavior of AgiWorld-WM
Camera and prompt practice
It extends what is already in the middle of the frame while the camera moves smoothly and most of what the previous seconds showed stays in view.
The camera turns gently while it walks, so that most of what the last five seconds showed stays in view. It walks in the direction it looks, and walks backward only for a few seconds.
A space that it has never shown is hard to generate at the moment the camera enters it. Walking through a doorway or around a corner onto unseen ground can produce a jump.
A new space is kept in the middle of the frame from the first chunk in which it is visible, and the camera walks toward it without stopping. The prompt of each of these chunks states what is there. The chunk that crosses an opening ends well beyond it.
Its memory holds only what it has generated. Parts of a room that were never in view have nothing to stay consistent with.
In a room, the first chunk turns the view across the room, about a quarter turn while walking slowly, so that its walls are in the memory before the camera leaves them. To look back, the camera turns around while walking.
It keeps a large space most consistent when the space is shown whole from a distance. Narrow views that each cover a small patch of a room give it little to keep consistent.
Large scenes are shown from far. An element that the instruction names is held in the middle of the frame, from far enough back that the space around it is also in view.
A wall or a shelf that fills the frame from close range gives it nothing to continue.
The camera looks where the space ahead opens, down a lane or across a room, rather than at the wall it passes. It moves sideways only while the view stays open.
It generates what the prompt names where the frame gives it room, and what it generates can differ from the layout.
The prompt names only what the frames of the chunk will certainly show. The next path is designed on the generated geometry.
Table 6: The specification of AgiWorld-WM shared by the initial design, the refresh, and the trajectory step. Each practice follows from the behavior on its left.
Check
Bound
Collision
A sphere of radius 0.25 m around the camera stays clear of the confirmed solids and of the point cloud.
Eye height
The eye stays within 0.7 to 1.4 times its starting height above the walked surface of the layout.
Support
The camera does not pass over an area that the layout marks as not walkable.
Speed
At most 2.5 m/s, with a change of pace of at most 0.6 m/s 2 .
Turning
At most 90 ∘ /s, and at most 25 ∘ /s on average over a chunk.
Chunk boundary
Between consecutive planned chunks, the velocity changes by at most 1 m/s and the turn rate by at most 45 ∘ /s.
Table 7: The main bounds on the solved path. A violation is returned to the director.