Interactive
Drift
Walk through a room while a model imagines it, one frame at a time. Watch small mistakes copy themselves. Then switch on the fixes researchers use.
A toy. No neural network runs here. A few rules stand in for a generator so you can see the shape of the failure.
Step 0
Room wrong 0%
The same 300-step walk, four ways.
What you are seeing
Close your eyes. Picture your bedroom. Now picture walking out to the kitchen and back.
Is your bed still where it was? For you, yes. Your brain keeps the room steady.
A computer that makes up a world as you walk has a harder job. It draws each new picture by looking at the last picture it drew. Not at the real room.
So one tiny mistake gets copied. Then the copy gets a mistake too. That is drift.
Press Walk and watch the right side. The purple boxes are the mistakes. See how they grow next to each other?
Now tick the memory notebook. The computer writes down what it saw. When it comes back, it checks the notebook instead of guessing.
An interactive video world model is autoregressive. Each frame is generated from the frames it generated before. During training it saw real frames. At run time it sees its own. That gap is exposure bias. The errors compound.
Two failures show up in long sessions:
- Compounding error. Small slips in view raise the chance of slips next to them.
- Lost permanence. Whatever leaves the view must be regenerated on return. Without memory it comes back different.
The toy has three switches. Each maps to a line of current research:
- Train on its own mistakes. Train on the model's own rollouts, not only clean frames. Diffusion Forcing and Self Forcing work this way. In the toy, slips are rarer and spread less.
- Memory notebook. Keep an explicit memory of what was seen, such as retrieved past frames or a 3D map. In the toy, unseen cells are frozen as last seen.
- Re-anchor every 25 steps. Ground the model in a real observation now and then, like a robot's camera. In the toy, the view resets to the real room.
Run the comparison. Training alone slows drift. Memory is what keeps the room. Grounding keeps it close to true.
This is a toy. No neural network runs here. About 200 lines of rules stand in for a generator. The numbers show the shape of the problem, not any real model's error rate. Google DeepMind names this same drift as an open limit of Genie 4. T-12 covers memory and consistency in depth.
Generated worlds are now good enough to walk through. The weak point is time. The longer a session runs, the more the world quietly changes.
That matters for any use that relies on the scene staying put. Training simulations, walkthroughs and agent testing all do.
Three questions for any vendor:
- How long does a scene stay consistent? Ask for minutes, not a demo clip.
- What happens when a user leaves an area and comes back?
- Can the system re-anchor to real data, such as a floor plan or live camera?
The comparison chart is the slide to take into the room. Training alone slows drift. Memory and grounding stop most of it. I would recommend to ask which of the three a product actually has.
Go deeper: T-11 Video generation, T-12 Interaction and memory and T-21 Evaluating world models.