Introduction

Most AI video tools still feel like a postcard: you prompt, you wait, you get a clip. JD.com's exploration lab just cracked the door the other way. At JDDiscovery-2026, the team open-sourced JoyAI-EchoWM (and a leaner EchoWM-Flash twin): omnimodal world models that keep painting the scene while you move, with ambient sound, music, and speech riding along.

Think less "render me eight seconds" and more "step inside, look around, listen." For creators chasing interactive media, that shift is sunshine after a long hallway of static demos.

What EchoWM Actually Ships

EchoWM is built for enterable generative media. Give it a first-frame image plus a world description, then steer with camera intent. The model answers in 720p video while generating environmental audio, music, and speech in the same pass.

Control is unified around a metric-scale 6-DoF trajectory:

  • In first person, the trajectory is your motion.
  • In third person, the model learns how subjects and framing should co-move — no separate follow-cam controller.

Keyboard or controller commands map into that same continuous pose stream, so roaming and follow shots share one language. A short trailing audiovisual window becomes context for the next segment, which helps layout and sound continue across multi-turn exploration (though the team is candid: there is still no full persistent 3D memory, so very long runs can drift).

EchoWM-Flash for Speed

EchoWM-Flash is the causal streaming sibling: a four-step distilled student aimed at interactive preview. Public weights list a Flash Preview with bounded sink/FIFO cache settings for autoregressive rollouts, with a longer-horizon Flash still on the roadmap. Base stays the quality-first bidirectional path (~10 s horizons in the current public build).

Why Creators Should Care

One-shot cinematic clips are great for ads and moodboards. EchoWM is aiming at something stickier:

  • Worlds you can revisit — geometry, subjects, and audio stay coherent as you turn and return
  • Native soundtrack — footsteps, collisions, dialogue, and score without a bolted-on audio model
  • Shared control surface — one Action DSL / camera path for both FP and TP scenes

On WBench Navigation, JoyAI-EchoWM landed around 81.6–81.7 average across reported cases, with consistency and interactivity among the stronger sub-scores. Flash stayed near 81.0 while keeping high interaction marks. Extra checks on SANA-WM-Bench and preference studies lean on visual quality and camera following under short and long horizons.

Training mixed gameplay logs, human play footage, Unreal pose-truth simulation, and internet video through staged audiovisual pretraining, action fine-tuning, joint fine-tuning, and autoregressive post-training. Public weights build on LTX-2.3; commercial use of those Lightricks-derived pieces still needs the usual license care.

Pro Tip

If you are prototyping interactive lookbooks or explorable mood worlds, start with a clean first frame and a spatial prompt ("sunlit arcade corridor, wet tile reflections, distant chatter") rather than a shot list. Steer with short camera intents first — Flash for iteration, Base when you want the prettier take.

Where to Grab It

JD framed the drop beside bigger physical-AI bets (a path toward a huge domestic GPU cluster, embodied video collection, logistics Superbrain 3.0). For ArtRealm readers, the gift is simpler: open specs for worlds you can enter, not just watch.

Conclusion

Clip generators still rule the feed. EchoWM is a bright hint of what comes next — media that walks with you, hums underfoot, and keeps the room coherent when you glance back. Open weights, open code, open invitation to play.

—Aurelia ♡