Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation

Visko Orbis 1.0, from Team Visko, proposes a Live Model for video generation: unlike offline systems such as OpenAI’s Sora and Google DeepMind’s Veo, it remains active, accepts prompt changes during generation, and applies them to future uncommitted chunks without restarting. A chunk-wise latent-flow generator supports text-to-video, image-to-video, video continuation, multilingual prompts, and hour-scale rollouts through bounded multi-scale memory that preserves recent detail while compressing older history. Event-aligned temporal captions, rollout-corrupted history training, guidance distillation, self-forcing distribution matching, GRPO reward optimization, and a latent world model target continuity, prompt alignment, and physically plausible motion. The serving stack combines KV-cache reuse, compiled and fused transformer execution, Ulysses-style sequence parallelism, progressive decoding, and a reference-aware video upscaler to deliver 4K video at 24 FPS, with prompt updates visible in under 1 second on average. On a 74-case benchmark containing one- to three-minute outputs, Orbis records a DOVER aesthetic score of 0.8101 and a technical score of 0.5572, while its long-form human Arena evaluation gives it the highest overall Elo of 1838 and temporal-stability Elo of 1940 among nine systems. The results suggest its main advantage is not only image quality, but maintaining a steerable, coherent stream as events and instructions change.