What it is
Open-source MoE video generator for embodied tasks with T2I, T2V, and TI2V pipelines, prompt rewriting, and refiner support. Claims ~3× faster inference and top open-source RBench score. Generate embodied-focused images and videos from text or a first frame, with a prompt rewriter, auto-negative prompts, and optional refiner across diffusers or SGLang backends.
Key results
- Claims ~3× faster inference from its MoE architecture
- Supports T2I, T2V, TI2V, and refinement in released checkpoints
- Top open-source RBench score: 0.620 avg as of Jul 9, 2026
Shows that large open MoE video models can target embodied world understanding while staying competitive on public benchmarks.