InclusionAI's Invoker-Render 1.0 Turns Rough 3D Scenes Into Streaming Video
InclusionAI released open weights for a 33B proxy-to-video renderer that converts streaming 3D layouts into photorealistic footage with locked camera and motion.
Takeaways
- InclusionAI released Invoker-Render 1.0, a 33B proxy-conditioned video renderer with open weights.
- Inputs are a structural proxy video, one reference image, and a text prompt.
- Pairs a bidirectional creation model with a causal four-step streaming model for long-horizon rendering.
- Accepts semantic maps, 3D boxes, subject masks, shaded geometry, and skeletons as proxies.
- Trained on roughly 200 hours of multi-granularity data plus self-generated rollouts.
- Needs 8 GPUs, 132.5 GB weights, and ships under Apache-2.0 plus non-commercial CC BY-NC 4.0; see the GitHub repo.
InclusionAI has released Invoker-Render 1.0, a 33-billion-parameter video model that converts a rough 3D scene stream into rendered video. The model is designed to continue generating frames as new proxy frames arrive, and its weights are available on Hugging Face.
Invoker-Render provides the rendering layer for Code World Model 2.0. An upstream engine retains control of physics, animation, camera movement, scene layout, and object state. The model receives a lightweight visual representation of that state and generates the detailed appearance, allowing existing simulators and game engines to preserve deterministic behavior.
Three inputs control each frame
Each inference request combines structural guidance, an appearance reference, and a text prompt:
| Input | Expected format | Purpose |
|---|---|---|
| Appearance reference | 1344 × 768 RGB image | Anchors the scene’s visual style, subjects, and overall appearance |
| Structural proxy | 24 FPS video | Controls camera position, composition, geometry, and motion |
| Language prompt | Text | Describes visual details that the proxy does not encode |
Structural proxies can contain semantic maps, sparse 3D boxes, subject masks, shaded geometry, or boxes augmented with skeletons. InclusionAI calls this range multi-granularity control. A coarse proxy can specify a character’s path and leave local movement to the model, while articulated geometry can guide individual limbs.
Four denoising steps sustain the stream
Invoker-Render supports continuous generation rather than a predetermined clip length. That requires the model to process proxy frames causally, using past context without waiting for future frames, while limiting visual drift across its own generated history.