Evoke is a 14B, three-step, CFG-free autoregressive world model for persistent interactive world generation. It creates a world that can remember camera movement, respond to changing instructions, and continue generating coherent content for hours.
The architecture decouples persistent world state from the denoiser and addresses that state through camera pose. A long-horizon interactive teacher transfers extended-session behavior to the few-step model, while a multi-term memory and world-state bank feed autoregressive diffusion transformer blocks.
Evoke is useful for interactive simulations, game-like environments, embodied-agent research, virtual worlds, and creative exploration. Users can navigate, turn, introduce new objects or events, and alter the world while generation is already running instead of restarting from a new prompt.
Get more likes & reach the top of search results by adding this button on your site!
Get top updates in AI to your inbox every weekend. It's free!