EVOKE 14B: Alaya Lab's Open 3-Step Interactive World Model

Overview

EVOKE ("Endless Interactive World with Bounded State and Long-Horizon Supervision") is the successor to AlayaWorld from Alaya Lab. The paper describes a three-step, CFG-free world model that reports state-of-the-art results on the WBench interactive benchmark while staying competitive on VBench-Long and VBench-2.0.

The distilled model generates 384 x 640 @ 24 fps and stays coherent over 30-second rollouts, producing 1.5 seconds of video every 2.11 seconds on a single H200. Weights are open on Hugging Face, with inference and training launchers in the GitHub repository.

Official EVOKE overview video

Three-Step Generation, Zero CFG

Most few-step models inherit their ceiling from the teacher used for distillation. EVOKE redesigns the teacher for long-horizon interactive generation instead of treating it as a fixed generator: chunk-wise grouping, distant-frame retrieval, and a linear-attention global state keep the teacher's memory and compute growing linearly with session length, which is what makes 30-second self-forced supervision affordable.

A distribution-matching objective, applied under self-forced rollouts, transfers those capabilities to a three-step student that uses no classifier-free guidance — one forward pass per step instead of two.

EVOKE inference pipeline: persistent world-state bank feeds the autoregressive few-step denoiser

Endless, Not Windowed

Interactive world models face a conflict: keeping history in the denoiser context or the key-value cache grows cost with session length, while low-latency interaction demands few-step generation. EVOKE resolves this by externalizing persistent world state:

• Write: a monocular depth model estimates depth for each emitted chunk under its known camera poses and unprojects it into a persistent point cloud.
• Read: the current camera pose addresses the bank directly; up to eight sources ranked by co-visibility are fused with a batched z-buffered scatter, returning a warped image plus a per-pixel visibility mask.
• Evict: an optional retention window, enabled explicitly for hour-scale runs.

Only view-relevant information is retrieved, so the denoiser context stays bounded however long the session runs — no trading session length for memory.

Change the World Mid-Flight

Per-chunk conditioning lets the prompt change while the rollout is running — no cut, no restart. Each schedule below switches at chunk 3 of 6 (213 frames, 8.9 s):

	scene	the prompt switches to
	frozen tundra, polar daylight	an aurora ignites across the whole sky
	persistent world navigation	camera moves and turns while the world keeps its spatial state

Conditioning Modes

Capability demo: navigating a generated world like a game (from the official project page)

Release Contents

The release ships four models plus the base components:

Every EVOKE directory is the parent of a transformer/ and loads with from_pretrained(path, subfolder="transformer"). The distilled models were trained on v2v conditioning alone, so i2v and t2v on them are zero-shot; only stage1_camera_control has all three modes in distribution. The VAE, text encoder, tokenizer, and scheduler in evoke-base come from the released Helios base, which traces them to Wan.

Availability

EVOKE does not have native ComfyUI support at launch. The official release provides Python inference launchers (scripts/inference/infer_post_distill.sh) for text-to-video, image-to-video, video-to-video, and mid-rollout segment modes, plus training launchers for each stage. The ViGeo depth backend is required for the world-state bank; Depth-Anything-3 is optional. Weights are on Hugging Face at AlayaLab/Evoke.