ATI: Any Trajectory Instruction for Controllable Video Generation
ATI by ByteDance Research is a trajectory-based motion control framework for Wan2.1. Unifies object, local, and camera movement control using any trajectory instructions.
ATI
Wan2.1Video GenerationMotion ControlTrajectory ControlOpen SourceAny Trajectory Instruction for controllable video generation by ByteDance Research. Built on Wan2.1, it unifies object, local region, and camera movement control through trajectory-based motion guidance.
| Developer | ByteDance Research |
| Architecture | Wan2.1-based with trajectory conditioning |
| Model Sizes | 14B (fp16/fp8/fp8_e5m2) |
| Capabilities | Object trajectory, region motion, camera trajectory control |
Overview
ATI (Any Trajectory Instruction) is a trajectory-based motion control framework from ByteDance Research that brings precise motion control to Wan2.1 video generation. It unifies three types of motion control in a single framework: object-level trajectory, local region movement, and camera motion. Users can specify arbitrary trajectories that the model follows when generating video content.
The framework supports image-to-video generation at 480P resolution and provides fine-grained control over how objects and the camera move throughout the generated sequence.
ComfyUI Integration
ATI-Wan2.1 models are available for ComfyUI through the Kijai WanVideoWrapper. The fp8 and fp8_e5m2 quantized variants enable efficient inference on consumer GPUs.
Available Sizes
| Size | Precision | Resolution |
|---|---|---|
| 14B | fp16 | 480P I2V |
| 14B | fp8 | 480P I2V |
| 14B | fp8_e5m2 | 480P I2V |
Resources
| Resource | Link |
|---|---|
| GitHub Repository | github.com/bytedance/ATI |
| Project Page | anytraj.github.io |
| Hugging Face (Original) | bytedance-research/ATI |
| Hugging Face (Kijai) | Kijai/WanVideo_comfy |
Guides and workflows related to this model series.