AlayaLab/Evoke

What it solves

EVOKE is a world model designed to generate high-quality, long-horizon videos that avoid the typical memory and quality degradation seen in windowed generation. It solves the problem of "memory ceilings" by using an external world state bank to store scene geometry, allowing for virtually endless rollouts without increasing the compute or memory requirements of the denoiser.

How it works

EVOKE utilizes a three-step, CFG-free (Classifier-Free Guidance) inference process to generate video chunks autoregressively. It maintains a persistent point cloud in a "world state bank" by estimating depth from emitted chunks and unprojecting them into the bank. When generating new frames, the model reads from this bank based on the current camera pose, retrieving and warping only the necessary visual information. This process is supported by a multi-tier memory system (long, mid, and short-term) and a specialized teacher-student distillation pipeline that enables efficient, few-step generation.

Who it’s for

This project is for AI researchers and developers working on world models, long-form video generation, and camera-controllable synthetic environments.

Highlights

• Endless Generation: Uses a camera-indexed world state bank to keep the context bounded regardless of session length.
• High Efficiency: Generates 1.5 seconds of video every 2.11 seconds on a single H200 GPU using only 3 steps and zero CFG.
• Mid-flight Re-prompting: Allows users to change the text prompt while a rollout is running without needing to restart or cut the video.
• Camera Control: Supports precise camera movement and rotation via a dedicated Director Web UI.
• Flexible Conditioning: Offers text-to-video (t2v), image-to-video (i2v), and video-to-video (v2v) modes.