Orbis changes a video’s direction without interrupting the stream

Visko is launching Orbis 1.0, a model that generates a continuous video stream, retains a compressed memory of the scene, and incorporates new instructions while the video is playing.

The scene is already unfolding when the user decides to change what happens next. A new instruction is submitted, the setting transforms, and the action moves in another direction without restarting the generation process. That is the premise of Orbis 1.0, Visko’s first commercial model for interactive long-form video.

Most video models produce a fixed result. The user submits an instruction or image, waits for processing, and receives a completed clip. Any modification generally requires a new generation or additional editing.

Orbis replaces this workflow with a continuous process. The video is generated and streamed in successive chunks. A new instruction can be sent while it is running, then applied to the next section that has not yet been produced.

Visko calls this category “Live Models.” The term does not simply describe a fast model. It refers to a system that maintains state between chunks, continues running over time, and accepts interventions without resetting the entire session.

The company’s use of the word “world” requires some clarification. In its currently accessible form, Orbis primarily delivers a video stream with an accompanying audio track. The public API does not provide 3D geometry, a complete map of the environment, or detailed object states comparable to those of a game engine.

Viewers can therefore observe and direct a generated world, but they do not yet receive a three-dimensional space that can be directly explored or modified. The demonstrations create the impression of a living environment, while the technical interface remains organized around a continuous sequence of images and sounds.

Orbis accepts either a text instruction or text accompanied by a reference image. The reference establishes the opening chunk and helps preserve the composition, subject, or overall appearance throughout the rest of the session.

The Orbis 1.0 technical report also mentions continuation from an existing video. This video-to-video capability is part of the model presented by Visko, but it does not yet appear in Reactor’s documented public interface, which primarily exposes text and image inputs.

New instructions never alter material that has already been streamed. They affect only future chunks. The system can therefore gradually turn a forest into a snowy environment, introduce a character, or redirect an action without deliberately inserting a hard cut.

The public stream is generated in groups of 33 frames, representing approximately 1.8 seconds. A new instruction is applied at the boundary of the following chunk. The response is not instantaneous in the strictest sense, although it remains short enough for live interaction.

Visko reports an average visible response time below one second in the configuration used for its study. With the current API, the delay experienced by a user may depend on when the request arrives, the connection, the selected resolution, and service demand.

Continuity relies on multiscale memory. Recent elements are retained in greater detail, while older sections are progressively compressed. When a segment leaves the active window, some of its information is consolidated into a fixed-capacity state.

This method prevents the cost of each new chunk from increasing with the video’s total duration. A one-hour session does not require the model to revisit every frame generated since it began.

Visko’s “persistent memory” is therefore not a perfect archive. The system attempts to retain characters, locations, style, movement, and important events, but it does not necessarily preserve every object or detail in the same way.

As a session continues, the model must implicitly determine what should be kept. Visko does not yet provide controls that let users lock a memory, inspect the retained state, or recover that memory for transfer into another session.

The phrase “unbounded-length generation” should be understood in the same way. The architecture does not require the video’s duration to be defined at launch, and the processing cost per chunk remains relatively stable. This does not guarantee that a stream can run forever without interruption, drift, or information loss.

The report primarily documents comparisons involving videos lasting one to three minutes, along with an experimental one-hour run. No public test has yet demonstrated that the same world can be maintained for several days or weeks, or recovered after a complete infrastructure interruption.

Orbis uses causal generation. Each chunk is produced from the existing state, available memory, active instruction, and any visual reference. The model cannot consult a future section that has not yet been created.

Visko considers this progression better suited to representing motion and cause-and-effect relationships. An object falls according to its previous position, then the next situation is produced from that new state.

The company does not present Orbis as a deterministic scientific simulator. Its movements are still generated from data and consistency scores. A visually credible result does not guarantee exact compliance with forces, masses, or material properties.

During post-training, another model evaluates several possible continuations according to their consistency with previous events. This stage favors movements and transformations considered more plausible, without providing the guarantees of simulation software built around explicit physical rules.

Visko’s website presents several potential uses, including interactive stories, voice narration, livestreams, virtual companions, concerts, combat scenes, product presentations, and robot training. Each demonstration follows the same basic principle: a specialized interface converts user actions into new conditions for the stream.

The narration feature, for example, lets users read a story aloud while the video is being generated. The documentation does not yet explain whether speech recognition is handled directly by Orbis or by a separate layer placed in front of