Papers.Data.Code
Repo

Evoke

AlayaLab

Generates endless interactive video by storing scene geometry in an external world-state bank and retrieving only the current view needs.

GitHub ↗Aug 12, 2026Paper ↗

  • cv
  • world-model
  • video-generation
  • camera-control
Interactive scene with Move/Rot controls showing rendering/navigation capability, likely a demo output
Interactive scene with Move/Rot controls showing rendering/navigation capability, likely a demo output

What it is

Three-step CFG-free world model for long interactive video with bounded memory, mid-rollout re-prompting, and SOTA on WBench. Generate long camera-controlled videos with bounded context, persistent scene memory, and prompt changes during rollout. Supports t2v, i2v, and v2v inference in 3 CFG-free steps.

Key results

  • 1.5 s of 384×640 video every 2.11 s on one H200
  • Runs in 3 steps with zero CFG; one forward pass per step
  • Supports mid-rollout prompt switches over 213-frame, 8.9 s demos

Shows long-horizon interactive video can stay bounded in context by moving scene memory outside the denoiser.

One useful ML email a week.

More from Aug 17 – Aug 23

The whole week →

Picked Aug 21

Picked Aug 20

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.