Adobe's Wonder: a minute-long video world you steer at 16 FPS
An Adobe–Johns Hopkins model holds a scene stable as you fly a 6-DoF camera through it in real time. No weights yet — but the recipe is public.
Copy markdown
A world you explore, not a clip you watch
Feed Wonder one image and it builds a navigable world you fly a camera through — new views render at 16 FPS with a constant ~0.5s latency and stay coherent for up to a minute. Turn around, come back, and the spot looks the same: it's a persistent scene, not a fresh hallucination every frame.
How it dodges the drift-and-forget problem
Two ideas carry it. A dense coordinate field injects camera pose as spatially-aligned pixels, so the model reads motion directly instead of inferring it. A sparse-attention memory then pulls only the relevant past tokens no matter how far you've wandered — the fix for the drift and forgetting that wrecks long video generation.
Two modes: image-to-video and re-shoot
It runs image-to-video (a still becomes a world you can walk) and video-to-video (hand it a clip, re-shoot it from a new camera path). That V2V mode is the quiet unlock: change the camera move on footage you already have, after the fact.
Why builders should care: this is game-engine energy
Persistent, camera-steerable worlds nudge generative video from clip-maker toward real-time engine. Think level blockouts from a single concept image, cheap simulation environments for training robots and agents, or virtual re-shoots for film and VFX — projects that needed a hand-built 3D scene yesterday.
The catch: weights are 'coming soon'
This is a research drop. The arXiv paper and demo reels are live, but code and weights are still marked 'coming soon,' and the write-up stays quiet on output resolution and GPU cost. Nothing to run tonight — but it's the world-model bar to watch, so star the repo.