AlayaLab/Evoke

Official implementation of EVOKE: Endless Interactive World with Bounded State and Long-Horizon Supervision. A three-step, CFG-free interactive world model. SOTA on WBench.

What it solves

EVOKE is a world model designed to generate high-quality, consistent, and "endless" video rollouts without the memory bottlenecks typically associated with long-horizon generation. It solves the problem of trading session length for memory by moving scene geometry into an external bank rather than relying solely on the denoiser's context window.

How it works

EVOKE uses a three-step, CFG-free (Classifier-Free Guidance) inference process. It maintains a "world state bank" that stores scene geometry as a persistent point cloud. During generation, the model reads from this bank based on the current camera pose to retrieve a warped image and visibility mask, which are then used to condition the generation of the next video chunk.

Key technical components include:

  • World State Bank: Handles writing (estimating depth via monocular models like ViGeo or Depth-Anything-3), reading (addressing the bank via camera pose), and eviction (managing retention windows).
  • Multi-tier Memory: Uses a combination of long, mid, and short-term parametric memory with varying patch kernels to maintain continuity.
  • Autoregressive Chunking: Generates video in latent chunks, using a global anchor and previous chunk latents to ensure stability.
  • Distillation: Employs a multi-stage distillation process to reduce the number of sampling steps to just three while maintaining visual quality.

Who it’s for

This project is for AI researchers and developers working on world models, autonomous agent simulation, and long-form video generation who need high-speed inference and consistent spatial memory over extended durations.

Highlights

  • High-Speed Inference: Generates 1.5 seconds of video every 2.11 seconds on a single H200 GPU using only 3 steps and zero CFG.
  • Endless Rollouts: Uses an external world state bank to keep the context bounded, allowing for theoretically infinite session lengths.
  • Mid-flight Re-prompting: Supports changing the text prompt while a rollout is running without needing to restart or cut the video.
  • Flexible Conditioning: Supports text-to-video (t2v), image-to-video (i2v), and video-to-video (v2v) modes.
  • Interactive Control: Includes a "Director Web UI" for real-time camera control, per-chunk prompting, and revision branching.

Related

  • Project
    thu-ml/Causal-Forcing

    A framework for high-quality real-time interactive video generation using Causal ODE and Consistency Distillation to optimize autoregressive diffusion distillation.

  • Project
    Junchao-cs/SolarWM

    SolarWM is an open‑source framework that provides a massive, camera‑annotated video dataset, a three‑stage training recipe, and ready‑to‑use 5 B–33 B video world‑model checkpoints (Wan2.2, LTX‑2.5, MiniMax‑H3). It lets you train and run long‑horizon interactive video models with a unified CLI and fully documented pipelines.

  • Project
    open-gigaai/giga-world-1

    An open-source framework for training and deploying robot world models to evaluate robot policies through controllable video generation.

  • Project
    microsoft/LatentSpatialMemory

    A framework for video world models that stores 3D scene content as latent tokens to improve generation speed and reduce memory overhead by avoiding repeated RGB rendering.

  • Project
    princeton-vl/infinigen

    A procedural generation engine that creates infinite photorealistic 3D worlds, indoor scenes, and articulated assets for computer vision and robotics simulation.