Looking Ahead Modeling Future Evolution in Autoregressive Video Generation paper:

Thanks AK for forwarding🤝

how long a clip can it hold together before things start to fall apart?

Planning for how a scene will evolve, not just the next frame, feels like the right direction. Keeping long videos consistent is still the hard part.

at what rollout length does the distribution shift between planning and acting frames collapse in autoregressive video - around 5-6s, or does the evolution model keep coherence past that? same sim-to-real gap we fight in robot control, just in pixel space.

Does the evolution model learn a continuous trajectory over frames, or does it still use discrete step scheduling that limits long coherent videos?

Thanks for sharing our work, AK! This is a proof-of-concept study, and our data collection and training pipeline are at an early stage. We'll keep exploring this direction and hope to share more interesting results. We'd love to hear feedback and suggestions from the community!

🙏

so the model finally stops improvising frame by frame, like me giving a standup update with zero notes

long clips always drift this could actually fix that

Planning how a scene will evolve, not just the next frame, is the shift that matters — autoregressive video's failure mode has always been drift over long clips. The interesting test is whether this holds coherence past the point where planning and acting frames diverge, especially where the same gap shows up in robot control.