Adobe and UT Austin's AgiWorld builds persistent video worlds

Adobe Research and the University of Texas at Austin posted a paper on November 2, 2026 that treats persistence as the unsolved problem in video world models. The paper, AgiWorld: An Agentic Video World Model for Building Persistent Worlds, describes a system that plans a world and a route through it, generates the video a few seconds at a time, and feeds what it generated back into the plan. The authors write that the world stays coherent because of that loop, and they report their benchmark results as the strongest so far on world revisiting and exploration.

Video world models are the branch of generative video that simulates a scene a camera can travel through, rather than a clip watched from one viewpoint. Persistence, the property that a place looks the same when the camera returns, has been a standing difficulty. The paper's related-work section records that autoregressive video models attend only to a bounded window of recent frames, so a region drifts in content, appearance and geometry once it leaves that window. Later methods widen the window without guaranteeing that a revisited region matches what came before.

Existing interactive world models take a single image, a camera path and a prompt, then generate video as the camera moves. The camera path stays the user's problem. A creator who wants to walk through a room, turn a corner and come back has to author every move, and the model decides on its own what sits beyond the edge of the frame. AgiWorld is an attempt to remove that manual step.

What AgiWorld is

The project page lists ten authors across two institutions. It credits Shaoheng Fang, Georgios Pavlakos and Qixing Huang to the University of Texas at Austin, and Yiqun Mei, Hanwen Jiang, Xin Sun, Haitian Zheng, Hao Tan, Kalyan Sunkavalli and Yiwei Hu to Adobe Research.

AgiWorld is two components joined at a shared memory. AgiWorld-Agent is a planner built on a multimodal language model. It keeps a world state that records the scene's layout in two parts: confirmed elements that have already appeared in the video, and imagined elements the agent has designed but the camera has not reached. A director inside the agent turns the next stage of the journey into a camera move, such as walking through a door or returning to a place seen before, and writes it as a metric camera trajectory and a structured text prompt.

AgiWorld-WM is the generator. It is a camera-conditioned autoregressive video model that produces the video five seconds at a time, follows the director's trajectory through a camera adapter and a relative camera positional encoding the paper calls PRoPE, and stores each chunk alongside its estimated depth in a point-cloud spatial memory.

After every chunk, the agent reads that memory, reconciles its plan with what was generated, confirms or retires the imagined elements, and extends the design ahead. The abstract calls the arrangement a closed loop and says it keeps the world faithful to the instruction and the agent's world state consistent with the generated content.

The two memory techniques

The memory carries the paper's technical contribution. Reusing stored geometry over a long rollout raises two problems: a past view that overlaps the new camera may show surfaces the new camera cannot see, and the estimated depth carries error that compounds as one chunk builds on the next.

Occlusion-aware history selection answers the first. It retrieves past views by the surfaces visible from the new camera, rather than by camera overlap alone. Reliability-based memory maintenance answers the second. It prefers geometry the depth estimator scored as confident, surfaces observed close to head-on, and points near the observing camera, and it favors earlier observations of each surface. Unreliable points are pushed down the order but never discarded, so a surface the system does not trust still hides what lies behind it.

The paper says the agent and the world model were co-designed around the shared spatial memory and around the interface of metric camera trajectories and structured prompts that passes between them.

What the benchmarks show

The authors evaluate the world model on two 60-second benchmarks, and they report persistent worlds built from free-form instructions over minute-long journeys. DL3DV-Evaluation is the world-revisiting test: its camera trajectories circle a scene and finish where they started, so any revisited region has to match what the model generated earlier. The authors also introduce a benchmark called World Explore, made of forty held-out outdoor scenes in which the camera keeps moving into parts of the scene the input image does not show.

On both, the paper reports the model ahead of seven named baselines, among them Lyra 2.0, Matrix-Game 3.0, LingBot-World, GEN3C and SANA-WM. The authors report cutting FID, a distribution-distance score where lower is better, from 25.54 to 21.74 on DL3DV-Evaluation and from 42.73 to 35.37 on World Explore. They report cutting the translation error on World Explore from the strongest baseline's 1.32 meters to 1.05 meters. An ablation finds both memory mechanisms carry weight, with FID over the last ten seconds rising from 51.84 to 57.74 when both are removed.

The results are the authors' own, measured on benchmarks that include one they designed themselves. AgiWorld is also an arXiv preprint and has not been peer-reviewed.

What it means for creators

Nothing in the paper is runnable today. The project page links the paper PDF, the arXiv listing and the authors' personal pages, and hosts the system's result videos, but it lists no code repository, no released weights and no public demo.

The idea worth watching is the split. If persistence comes from an agent loop around a generator, then progress on explorable worlds depends on planning and memory as much as on the next video model. For creators, that points at previsualization, virtual sets and interactive pieces, where the manual work today is laying out every camera move. AgiWorld does not remove that work yet, and whether an agent loop of this kind runs on the compute a small studio has is a question the paper leaves open.

The paper is on arXiv, with result videos on the project page; no code, weights or demo have been published.