Orbis Operational

Orbis is Visko's first Live Model, enabling AI video to run continuously like a livestream. Users can enter new instructions at any time to change the direction of the scene in real time, without waiting for a finished video. The model supports 4K@24FPS real-time streaming generation, maintains hour-scale long-term consistency through a multi-scale memory mechanism, and ensures realistic, natural motion through causal temporal decomposition and physics-aware rewards. It is suitable for virtual livestreaming, interactive storytelling, robot simulation, and other scenarios.

Real-time streaming generation: Video output continues uninterrupted like a livestream, supporting real-time delivery at 4K@24FPS.
Interaction during generation: Users can add or modify prompts at any time; new instructions affect subsequent frames within an average of 1 second.
Multimodal input: Supports text-to-video, image-to-video, and video continuation as three ways to start a world.
Long-term consistency: The multi-scale memory mechanism supports continuous generation for hours without noticeable quality degradation.
Physical realism: Causal decomposition and world-model rewards ensure that object motion and interactions follow the laws of physics.

Causal temporal decomposition: Generation is treated as an initial-value problem. Each chunk predicts only the next time segment from the committed history, naturally matching the Markov structure of physical evolution and avoiding the problem of “future noise contaminating current decisions” in bidirectional denoising.
Bounded multi-scale memory: Recent history is retained at its native latent resolution, while older history is progressively compressed. Information beyond the window is consolidated into a fixed-capacity learned state, decoupling memory and compute costs from rollout length.
Progressive training pipeline: A bidirectional model learns short-range visual priors, which are then distilled into a streaming generator. Event-aligned long-video training and human-curated fine-tuning follow, culminating in visual and physics-reward alignment through self-forcing distribution matching and GRPO reinforcement learning.
Physics-aware reward alignment: A frozen V-JEPA 2 latent world model evaluates the predictive consistency between candidate frames and historical states, guiding generation toward real-world dynamics and jointly optimizing it with visual-quality rewards.
Streaming inference engineering: State reuse, compiled Transformer execution, multi-GPU sequence parallelism, and a progressive decode-super-resolution-delivery pipeline work together to achieve real-time 4K delivery.

Start a session: Visit the Playground on the Visko website or call the API. Enter text, upload an image, or paste a video prefix to start a real-time session.
Start real-time generation: Click the Live button to start streaming generation. The scene will continuously evolve automatically like a livestream.
Intervene in the scene in real time: Enter new text prompts or voice instructions at any time during generation. Subsequent scenes can be changed in real time within an average of 1 second.
Switch scenes seamlessly: Describe a scene change with “Morph the scene.” The new instruction will be seamlessly incorporated at the next chunk boundary.
Choose output quality: Select 1080p, 2K, or 4K output resolution as needed. The system delivers high-quality video in real time through streaming super-resolution.
Capture and save clips: Click the Clip button at any time to capture and save the currently generated video segment for later editing or sharing.
Integrate with your own applications: Obtain an API Key to integrate Orbis's real-time generation capabilities into your own applications, livestreams, or robot simulation workflows.

Real-time streaming interaction: Video is continuously generated like a livestream, allowing users to enter new instructions at any time to change its direction in real time without waiting for a finished video.
Extended temporal consistency: The bounded multi-scale memory mechanism supports continuous generation for hours without noticeable quality degradation or color drift.
Strong physical realism: Causal temporal decomposition and latent world-model rewards ensure that object motion and interactions follow real-world physics.
High-quality real-time delivery: Native generation combined with streaming super-resolution enables real-time output of 4K@24FPS video.
Unified generation interface: The same causal continuation framework supports text-to-video, image-to-video, and video continuation as three starting methods.

Official website: https://www.visko.ai/models
Technical paper: https://www.visko.ai/blog/papers/files/Visko_Orbis_1_0__A_Live_Model_for_Real_TimeInteractive_Long_Video_Generation.pdf

Interactive entertainment and storytelling: Users choose story branches in real time or narrate stories by voice, while the scene evolves instantly in response to their instructions, creating a “playable movie” experience.
AI livestreaming and virtual hosts: As a real-time visual engine for virtual hosts, it continuously generates livestream scenes and responds to comments or voice instructions without prerecorded footage.
Robot simulation training: Train and test robot behavior in generated, high-fidelity virtual worlds, reducing the cost of trial and error before deployment in real environments.
Immersive education and training: Visualize abstract knowledge or historical scenes in real time, allowing students to instantly change demonstrations through voice questions.
Virtual e-commerce presentations: A virtual influencer dynamically showcases products in real-time generated scenes, while users can instantly request changes to the style, angle, or usage scenario.