ScenesAI demos real-time shapeshifting on RTX 5090
DAWA and Cyanpuppets showcased ScenesAI running interactive video-to-video transformations at 25–29 FPS on a single NVIDIA RTX 5090. The spatial world model tracks 3D motion and occlusion from one camera feed, enabling instant character and environment restyling with zero motion jitter.
Running interactive world models locally on consumer silicon marks a critical shift away from latency-heavy, credit-metered cloud video pipelines. By decoupling 3D kinematic tracking from selective neural rendering, ScenesAI delivers the temporal continuity that existing video diffusion models sorely lack.
- –Selective frame updates: The model recalculates only modified regions (around 20% of pixels) rather than regenerating entire frames, sustaining 25+ FPS live streaming through NVENC.
- –Persistent spatial memory: Scene mutations and character edits remain anchored in the 3D environment across frames, preventing the identity drift and morphing common in generative video.
- –Consumer hardware viability: Executing entirely on a single RTX 5090 brings real-time virtual production and VTuber avatars to indie creators without multi-GPU clusters.
- –Kinematic grounding: Leveraging Cyanpuppets' markerless motion capture gives the model explicit physical constraints, maintaining natural actor momentum and occlusion boundaries.
DISCOVERED
11m ago
2026-11-16
PUBLISHED
39m ago
2026-11-16
RELEVANCE
AUTHOR
cyanpuppets
