# AnySplat — card for AI agents

Feed-forward model that turns unconstrained multi-view images into 3D Gaussian splats, with open code and weights.

Category: AI 3D model generators · maker Lihan Jiang et al. (OpenRobotLab) · checked 2026-11-01

## Price
- Price: Free
- Free tier: Free open-source code and model weights under the MIT License

## License
MIT License — https://huggingface.co/lhjiang/anysplat

- commercial: **yes** (Commercial use OK)
  > The code and models are licensed under the MIT License
  (https://huggingface.co/lhjiang/anysplat)
- attribution: **no** (No credit needed)
  > The code and models are licensed under the MIT License
  (https://huggingface.co/lhjiang/anysplat)
- in app: **yes** (Can ship inside your app)
  > The code and models are licensed under the MIT License
  (https://huggingface.co/lhjiang/anysplat)

## Rules to follow
- Keep the MIT license notice with any code or model weights you ship.
- Install and run via the GitHub repository (OpenRobotLab/AnySplat) following its instructions.
- Cite the AnySplat paper (arXiv:2505.23716) when publishing results.

## How to connect
- Download only
- Repository: https://github.com/InternRobotics/AnySplat
- Website: https://huggingface.co/lhjiang/anysplat

## Notes
Added automatically; license read from the official pages.

## Numbers (AgentGid data)
- Gid Score: 7
- GitHub stars: 970

## Lines to paste into your prompt or AGENTS.md

```
- AnySplat (AI 3D model generators). Feed-forward model that turns unconstrained multi-view images into 3D Gaussian splats, with open code and weights. Price: free; commercial use OK; no credit needed.
  Connect: download from https://huggingface.co/lhjiang/anysplat (no API).
  Rule: Keep the MIT license notice with any code or model weights you ship.
  Rule: Install and run via the GitHub repository (OpenRobotLab/AnySplat) following its instructions.
  Rule: Cite the AnySplat paper (arXiv:2505.23716) when publishing results.
  License: https://huggingface.co/lhjiang/anysplat (checked 2026-11-01; details: https://agentgid.com/assets/anysplat/)
```

## Links
- AgentGid page: https://agentgid.com/assets/anysplat/
- Is it free for commercial use? https://agentgid.com/assets/anysplat/commercial-use/
- Source: https://huggingface.co/lhjiang/anysplat


## AnySplat: a 3D scene from a few ordinary photos

AnySplat is a free research model that turns a handful of ordinary photos or a short video of a place into a 3D scene you can look around in. It works out where each photo was taken by itself, in one quick pass. It needs a computer with an NVIDIA graphics card, and the code and model are MIT licensed.

### Where to use it
- A 3D view of a room from phone photos: Take a few photos of a room, a shop or a garden from different spots and get a 3D scene you can turn and move through.
- A 3D scene from a short video: Walk around a place with your phone camera, upload the video, and get the scene in 3D with a flythrough video.
- A scene file for a web viewer: Save the result as a .ply splat file and open it in a separate splat viewer or editor, for example to clean it up.
- Camera positions for other tools: Get the position and lens settings of every photo, which other 3D and mapping tools need as a starting point.
- Training on your own scenes: Researchers can train the model further on their own collections of scenes, using the included dataset templates.
- A 3D preview for a listing page: Turn a few photos of a flat, a venue or a product into a scene visitors can turn around on a page, next to the usual photos. (idea, not from the official pages)

### How to start
- Get the code and make a Python 3.10 environment with PyTorch for CUDA 12.1 on a Linux machine with an NVIDIA card.

```
git clone https://github.com/OpenRobotLab/AnySplat.git
cd AnySplat
conda create -y -n anysplat python=3.10
conda activate anysplat
pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
```

- Run the example script on a folder of photos; the model downloads from Hugging Face on first use.

```
python inference.py
```


### License
MIT license for the code and the model: free to use, change and share, also in paid products. Keep the copyright and license notice with copies; the authors ask you to cite their paper in published work.

### Good to know
- It needs a Linux computer with an NVIDIA card; one required part is only published for Linux, Python 3.10 and CUDA 12.1.
- The model file is about 2.9 GB and downloads on the first run.
- The ready demo page shares itself through a public link by default, so anyone with the link could use it while it runs.
- The resulting splat files are large; the authors note they may not load in the online demo and suggest other viewers.
- It is research code from a university and lab team, released in June 2025, not a supported product.

Checked 2026-11-19 against: https://huggingface.co/lhjiang/anysplat, https://huggingface.co/lhjiang/anysplat/raw/main/README.md, https://huggingface.co/api/models/lhjiang/anysplat, https://huggingface.co/api/models/lhjiang/anysplat/tree/main, https://github.com/InternRobotics/AnySplat, https://github.com/InternRobotics/AnySplat/blob/main/README.md, https://github.com/InternRobotics/AnySplat/blob/main/LICENSE, https://github.com/InternRobotics/AnySplat/blob/main/requirements.txt, https://github.com/InternRobotics/AnySplat/blob/main/inference.py, https://github.com/InternRobotics/AnySplat/blob/main/demo_gradio.py, https://github.com/InternRobotics/AnySplat/blob/main/config/main.yaml, https://city-super.github.io/anysplat/, https://arxiv.org/abs/2505.23716

### Mistakes to avoid
- Don't: AnySplat.from_pretrained("anysplat_ckpt_v1"), as in the Hugging Face model card → Do: AnySplat.from_pretrained("lhjiang/anysplat"), as in the GitHub README and inference.py (The model card snippet uses a local folder name; the Hugging Face model id is lhjiang/anysplat.)
- Don't: Running python demo_gradio.py on private photos as it is → Do: Change share=True to share=False in the launch line so the page stays on your machine (The script asks Gradio for a public link by default.)
- Don't: Starting training with the default config → Do: Set wandb.mode=disabled (or your own Weights & Biases project) on the command line (The default config logs online to the authors' team project.)
- Don't: pip install -r requirements.txt on macOS, Windows or Python 3.11+ → Do: Use Linux x86_64 with Python 3.10 and CUDA 12.1, or build gsplat yourself (requirements.txt pins a gsplat wheel built only for that combination.)
- Don't: Running inference.py from another folder → Do: Run it from the AnySplat folder, since it imports from the src package by path (The script adds the parent folder to sys.path and imports src.model, which only works next to the code.)

### Questions
- Do I need to know where each photo was taken? No. The model works out the camera position and lens of each photo itself.
- Can I use it on a Mac? Not as published. The install pins a Linux-only build for NVIDIA cards with CUDA 12.1 and Python 3.10.
- Can I use it in a commercial product? The code and model are MIT licensed, which allows commercial use with the license notice kept.
- What do I get out of it? A .ply file with the 3D splats, plus a rendered color video and a depth video along a path between your photos.
- What size are the input photos? The example code scales each photo to 448 by 448 pixels before running the model.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
