Invoker-Render is the renderer for Code World Model 2.0 (Ant Group inclusionAI, Westlake University, Zhejiang University): it takes a structural proxy video (semantic maps, sparse 3D boxes, subject masks, shaded geometry or skeletonized boxes), an appearance reference image, and text as input, generating video in a causal streaming manner with chunks of 4 steps, where each chunk only attends to fixed appearance anchors and limited recent history. The weights are for the main model with 33.12B parameters in FP32 precision (approximately 132.5 GB), following the MiniMax-H3 architecture. The inference code is based on SGLang and SGLang-Omni, featuring a WebSocket session service and a browser client.
Why it matters Camera, layout, and object motion are fixed by the upstream engine's wireframes or semantic maps; the model only renders materials and lighting, enabling continuous streaming generation. The authors demonstrated thirty minutes of continuous racing footage.
Hugging Face Likes 8 · Hugging Face Trending Score 8 · Hugging Face Downloads 1989
License: Non-commercial · ComfyUI: Not available · VRAM: ≈80 GB estimated
Reasons
- Includes paper, project page, inference code, and online demo on Hugging Face
- Authors claim it ranks first in all four human evaluation metrics on the 140 held-out test segments
Usage
- Prepare the environment and inference package according to the README and docs/INFERENCE.md at github.com/inclusionAI/CodeWorldModel. Input a 1344×768 first-frame image, a 24 FPS proxy video, and text, then run invoker_render.py for file-based inference or start the session service.
Risks
- The default production configuration requires a single machine with 8 NVIDIA GPUs; the repository does not include the VAE, Qwen, causal decoder, etc., which must be prepared separately.
- The license is Apache-2.0 plus CC BY-NC 4.0, treated as non-commercial; training methods and data have not yet been disclosed.