--- title: inspatio-world-v1.5 canonical_url: "https://www.modelscope.ai/models/InSpatio/inspatio-world-v1.5" md_url: "https://www.modelscope.ai/models/InSpatio/inspatio-world-v1.5.md" repository: InSpatio/inspatio-world-v1.5 last_updated: 2026-09-30 pipeline_tag: any-to-any tasks: - any-to-any parameters: 1.4B tensor_type: - F32 library_name: - safetensors downloads: 8 stars: 0 --- # inspatio-world-v1.5 > inspatio-world-v1.5 - An open-source model by InSpatio on ModelScope. 🤖 ModelScope | 🤗 HuggingFace | 💻 GitHub | 🌐 Project Page | 📑 Paper |… - **Repository**: InSpatio/inspatio-world-v1.5 - **Tasks**: any-to-any - **Parameters**: 1.4B - **Downloads**: 8 - **Stars**: 0 - **Last updated**: 2026-09-30 Source: https://www.modelscope.ai/models/InSpatio/inspatio-world-v1.5 ---

InSpatio-World-1.5

## Introduction **InSpatio-World** is a real-time 4D world simulator via spatiotemporal autoregressive modeling. This repository provides the **InSpatio-World-1.5 1.3B checkpoint**. - **Multiple input modes** — Run inference from a single image, four images, or a video. - **Target camera trajectories** — Supply a target camera trajectory to generate video from the requested viewpoints. - **Depth and source camera estimation** — For video inputs, Depth Anything 3 estimates missing depth and per-frame source camera intrinsics and extrinsics. - **Real-time examples** — See the single-image and multi-image inference results in the showcase below. For more details, see the [GitHub repo](https://github.com/inspatio/inspatio-world-v1.5), [project page](https://inspatio.github.io/inspatio-world-v1.5/), and [paper](https://arxiv.org/abs/2604.07209). ## Quick Start ### Installation Requirements: **Python 3.10** and **CUDA 12.6**. ```bash git clone https://github.com/inspatio/inspatio-world-v1.5.git cd inspatio-world-v1.5 conda env create -f environment.yml conda activate inspatio_world_test python -m pip install --no-deps depth-anything-3==0.1.1 python -m pip install modelscope ``` The official `environment.yml` includes the packages DA3 uses on this inference path. Installing DA3 separately with `--no-deps` avoids installing its unrelated app and benchmark dependencies. On Hopper GPUs, the official repository recommends FlashAttention-3 (FA3) for faster attention. ### Run the Demo ```bash bash run_example.sh ``` This runs the six scenes listed in `examples/manifest.json`. Results are saved to: ```text output// ├── source.mp4 ├── render.mp4 ├── mask.mp4 └── pred.mp4 ``` Image and multi-image scenes reuse uint16 depth PNGs. For video inputs, DA3 estimates missing depth and per-frame source camera intrinsics and extrinsics. The target trajectory is supplied by the user. The default depth mode is `auto`; use `--depth_mode existing` to require existing depth, or `--depth_mode estimate` to regenerate it. ### Single-Image and Multi-Image Inputs Prepare a scene using the official [image and multi-image examples](https://github.com/inspatio/inspatio-world-v1.5/blob/main/examples/README.md), then run: ```bash bash run_inference.sh --scene_dir /path/to/my_scene ``` Inputs must be **832×480**. A scene includes RGB input, depth maps, a prompt, source camera intrinsics and poses, a target camera trajectory, and scene metadata. See the linked examples for the file layout and camera matrix conventions. ### Video Inputs Provide the RGB video, prompt text, and one target OpenCV world-to-camera **4×4 matrix per frame** (16 row-major numbers per line): ```bash bash run_inference.sh --video /path/to/input.mp4 \ --prompt "Your prompt" --target_traj /path/to/target_tcw.txt ``` The runner reads the video's frame rate and frame count, resizes it to **832×480** if needed, and estimates depth and per-frame source cameras with DA3. The target trajectory must use the same first-frame-normalized coordinate system and displacement scale as the estimated source cameras. ## Showcase

Real-time inference results with a single input image

Real-time inference results: single-image input (left) and four-image input (right).

Try the [live demo](https://world.inspatio.com/) for an interactive experience. ## License The official repository's code is licensed under the [Apache-2.0 License](https://github.com/inspatio/inspatio-world-v1.5/blob/main/LICENSE). The upstream README states that this license applies only to code in its library; dependencies such as [Depth Anything 3](https://github.com/ByteDance-Seed/depth-anything-3) are separately licensed. ## Citation If you use InSpatio-World in your research, please use the BibTeX entry provided by the official repository: ```bibtex @misc{inspatio-world, title={INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling}, author={InSpatio Team}, journal={arXiv preprint arXiv: 2604.07209}, year={2026} } ``` ## Acknowledgement InSpatio-World uses a backbone based on [Wan2.1](https://github.com/Wan-Video/Wan2.1), with its training code referencing [Self-Forcing](https://github.com/guandeh17/Self-Forcing). We thank the Self-Forcing and Wan teams for their work and open-source contributions, and acknowledge [Depth Anything 3](https://github.com/ByteDance-Seed/depth-anything-3) and [ReCamMaster](https://github.com/KlingAIResearch/ReCamMaster) for their contributions.