Description
Invoker-Render is a sophisticated video-to-video AI model, developed by inclusionAI, that enables highly controllable video generation. With 33.12 billion parameters, this model focuses on generating video appearances based on structural proxies, a single appearance reference image, and textual prompts. It operates on the principle of keeping the world state in an upstream engine, where visual proxies dictate elements like camera path, scene layout, and object motion, while the video model handles the visual rendering.
Read More
The core of Invoker-Render lies in its causal streaming DiT weights, which are hosted in its repository. The model is designed to work in conjunction with the companion CodeWorldModel inference code, facilitating causal four-step streaming generation. This setup is built upon robust frameworks such as MiniMax-H3, SGLang, and SGLang-Omni, ensuring efficient and scalable operation.
Key highlights of Invoker-Render include its capability for long-horizon streaming, which allows sessions to run indefinitely as long as proxy frames are provided. This is achieved through a fixed appearance anchor and a bounded history window, maintaining visual consistency over extended periods. Furthermore, the model offers multi-granularity control, accepting a wide range of structural inputs such as semantic maps, sparse 3D boxes, subject masks, shaded geometry, and boxes with skeletons, providing users with precise control over the generated video content.
For interactive deployment, the companion code provides a WebSocket session service, a file client, and a browser client, making the model accessible for various applications. The model's weights are stored in FP32 format across 29 safetensors shards, totaling approximately 132.5 GB. While the repository includes the main causal streaming DiT weights, users need to prepare matching assets like VAE, Qwen, the causal decoder, and silence tensors separately for inference. The default production profile for inference typically requires eight NVIDIA GPUs on a single node, processing inputs like a 1344×768 RGB first image, a 24 FPS proxy video, and a text prompt.
- CategoryAI Models & LLMs
- DeploymentCloud
- PlatformsWeb
Invoker-Render Highlights
Video-to-Video Model: Generates video appearances from structural proxies and inputs.
33 Billion Parameters: Large model size for advanced video generation capabilities.
Proxy-Conditioned Rendering: Uses visual proxies to control camera path, scene layout, and object motion.
Appearance Reference Image Input: Takes one image to guide the visual style.
Text Prompt Input: Accepts textual descriptions for content generation.
Causal Streaming DiT Weights: Utilizes specific model weights for streaming generation.
Long-Horizon Streaming: Supports continuous video generation as proxy frames arrive.
Fixed Appearance Anchor: Maintains consistent appearance over long sessions.
Multi-Granularity Control: Accepts diverse structural inputs like semantic maps, 3D boxes, and subject masks.
Interactive Deployment: Companion code offers WebSocket, file, and browser clients.
FP32 Weights: Model weights stored in full precision.
MiniMax-H3 Base Model: Built on the MiniMax-H3 architecture.
Apache-2.0 License: Governs the repository's use.
CC BY-NC 4.0 License: Applies to non-commercial use.
Getting Started with Invoker-Render
Access Model Weights: Download the main causal streaming DiT weights from the repository root.
Prepare Assets: Separately acquire and prepare matching assets such as VAE, Qwen, and the causal decoder.
Set Up Environment: Follow the README.md and docs/INFERENCE.md in the companion CodeWorldModel repository.
Prepare Bundle: Create an inference bundle as instructed in the companion repository.
Start Service or Run Inference: Initiate the WebSocket session service or execute file inference using invoker_render.py.
Provide Inputs: Supply a 1344×768 RGB first image, a 24 FPS proxy video, and a text prompt.
Invoker-Render's Use Cases
- Controllable Video Generation
- Long-Form Video Creation
- Semantic Video Editing
- 3D Scene Rendering
- Interactive Video Applications




