EchoWM earns its omnimodal label by generating audio natively, not bolting it on

EchoWM generates 720p video, environmental sound, music, and speech simultaneously, and lets a user navigate through the result in real time. That combination — native audio-visual output plus continuous geometric control — is the gap it is designed to close. Most existing world models produce silent visual rollouts or lock navigation to a single domain's action vocabulary. EchoWM instead organizes everything around camera intent: a relative 6-DoF trajectory that means "move the observer this way," whether the scene is first-person or third-person. In third-person scenes, the model learns the camera–character coupling from data rather than from an explicit rig, so the same interface that steers an ego-view also orbits a character or tracks a vehicle.

The trajectory representation solves a concrete problem: heterogeneous sources — internally captured gameplay, human-played internet recordings, Unreal Engine renders, and general web video — all express motion at different scales. Per-clip normalization would erase displacement magnitude and introduce speed jumps at chunk boundaries. EchoWM instead estimates one robust scale from the 90th-percentile translation extent across the training set and applies it globally, preserving relative motion amplitude across sources and continuation turns.

Training is split into four stages that match each capability to the data where its supervision is cleanest. Audio-Visual Continued Pretraining runs on AV-rich clips to establish acoustic and visual priors. Action Fine-Tuning then freezes that backbone and trains only a lightweight camera-attention branch on control-clean trajectories, deliberately stripping the narrative annotation field so the model cannot infer motion from text. Joint Fine-Tuning updates both parameter groups at a reduced learning rate on a smaller balanced subset. A final autoregressive post-training stage — using Self-Gradient Forcing and Distribution Matching Distillation — converts the bidirectional diffusion model into a causal few-step generator stable under its own accumulated history.

On WBench Navigation's 158-case split, the undistilled model scores 81.7 average; the four-step causal variant EchoWM-Flash scores 81.0 while actually improving the Interaction dimension to 87.9. On SANA-WM-Bench's 961-frame long-horizon protocol, EchoWM-Flash achieves the highest VBench Overall on both Simple and Hard splits among causal methods, with stronger revisit consistency than competing causal models. The honest caveat the paper states: rotation error grows from the short-horizon figure to a larger value on Hard trajectories at 961 frames, and accumulated pose drift remains the main unsolved challenge.

The model explicitly does not handle arbitrary actor intent — jumping, attacking, manipulation — and has no persistent 3D memory, so geometry and identity can drift over long continuation. Those are real limits, not hedges.