Adobe's Wonder: Real-Time Interactive Video World Model with Constant Latency
aigclink · x · 2026-07-31
Adobe and Johns Hopkins University introduced Wonder, a real-time interactive video world model. It generates persistent, explorable worlds from an image or video, supporting 6-DoF camera control at a constant 0.5s latency with over a minute of consistency.
Core Breakthroughs
- Camera Control: Replaces numerical pose inputs with a 3D scaffold and environment map, rendering translation and rotation into frame-aligned visual evidence so the model "sees" camera motion.
- Long-term Memory: Maintains full-fidelity historical KV caches with sparse attention for on-demand retrieval, preventing scene drift when revisiting areas while keeping compute constant.
- Distillation Pipeline: Uses a Mixture-of-Students design to improve capacity and GAN regularization to resolve camera drift, preserving the teacher model's capabilities.
The model's V2V capability allows real-time re-capturing of existing videos from new angles, directly applicable to film pre-visualization, game prototyping, and robotics simulation. Code and weights are coming soon.
More from Multimodal
- Nokia vs iPhone, car edition: AI image shows Tesla Roadster hovering over a Porsche — vaibhavbetter · 2026-10-16
- Midjourney multi-sref weight mixing recipe for cinematic light & motion portraits — michaelrabone · 2026-10-16
- MiniCPM-V 4.7 35B-A3B hits 17+ tokens/s on iPad with ~3GB memory — techNmak · 2026-10-16
- turbopixel: layer-and-mask generative compositing runs a 9B model on 4GB GPU — 3unjee · 2026-10-16
- Manga Screen: open-source ComfyUI nodes turn illustrations into printed manga panels — KnowledgeBig4364 · 2026-10-16
- AI re-cut of Game of Thrones' Battle of the Bastards stars an orange cat — AxonkaiLab · 2026-10-16