Ant Group's Inclusion AI Unveils Invoker-Render for Stable 30-Minute Streaming World Models
inclusionAI · hf · 2026-11-19
Inclusion AI (Ant Group), Westlake University and Zhejiang University present Code World Model 2.0 and Invoker-Render, decoupling world-state simulation from photorealistic video synthesis to support both offline clip creation and real-time interactive streaming.
- Architecture: a bidirectional creation model paired with a causal four-step streaming model under a shared conditioning interface over structural proxies, an appearance reference, and language, trained on 200 hours of aligned multi-granularity data; it honors fine geometric constraints when given and synthesizes plausible motion from coarse proxies.
- ProForcing: trains a causal predictor, distills it into a four-step renderer using training-video histories, then shifts to continuous self-generated rollouts to prevent drift compounding.
- A pipelined serving stack enables efficient streaming inference, with stable 30-minute racing rollouts that preserve scene and subject appearance.
Related event: Ant's Inclusion AI Releases Code World Model 2.0 with Invoker-Render(2 posts)→
More from Multimodal
- Remotion ships real motion blur via HTML-in-canvas and WebGPU — Vjeux · 2026-11-19
- Blogger: Gemini 4 Nano's multimodal and design fidelity can replace Opus 5.5 in real tasks — op7418 · 2026-11-19
- StableCraft: open-source lasso inpainting linking ComfyUI to a Photoshop-style editor — Business_Plum4088 · 2026-11-19
- Krea ships a new UI: simpler, more powerful and much faster — TitusTeatus · 2026-11-19
- How are those AI face-replacement TikTok videos made? Redditor asks — rdditiszionist · 2026-11-19
- Omokage: a procedurally generated story website in a single HTML file with Fable 5.2 — jfischoff · 2026-11-19