| 1 | <p align="center"> |
| 2 | <img src="../FilmAgent-pics/cover_page.jpg" width="300" style="margin-bottom: 0.2;"/> |
| 3 | <p> |
| 4 | |
| 5 | <h2 align="center"> <a href="https://github.com/HITsz-TMG/FilmAgent">FilmAgent: A Multi-Agent Framework for End-to-End Film Automation in Virtual 3D Spaces</a></h2> |
| 6 | <!-- <h5 align="center"> If you like our project, please consider giving us a star ⭐ on GitHub to stay updated with the latest developments. </h2> --> |
| 7 | <h4 align="center"> |
| 8 | |
| 9 | <div align="center"> |
| 10 | <img src="https://img.shields.io/badge/Version-1.0.0-blue.svg" alt="Version"> |
| 11 | <img src="https://img.shields.io/badge/License-Apache%202.0-green.svg" alt="License"> |
| 12 | <img src="https://img.shields.io/github/stars/HITsz-TMG/FilmAgent?color=yellow" alt="Stars"> |
| 13 | <img src="https://img.shields.io/github/issues/HITsz-TMG/FilmAgent?color=red" alt="Issues"> |
| 14 | <img src="https://img.shields.io/badge/python-3.8-purple.svg" alt="Python"> |
| 15 | |
| 16 | <!-- <img src="https://img.shields.io/github/stars/AIDC-AI/Marco-o1?color=yellow" alt="Stars"> --> |
| 17 | |
| 18 | <!-- [](https://filmagent.github.io/) |
| 19 | [](https://arxiv.org/abs/2501.12909) |
| 20 | [](https://www.youtube.com/watch?v=hTI-0777iHU) |
| 21 |  --> |
| 22 | |
| 23 | </h4> |
| 24 | |
| 25 | <div align="center"> |
| 26 | |
| 27 | <!-- **Affiliations:** --> |
| 28 | |
| 29 | _**Zhenran Xu, Longyue Wang, Jifang Wang, Zhouyi Li, Senbao Shi, Xue Yang, Yiyu Wang, Baotian Hu, Jun Yu, Min Zhang**_ |
| 30 | |
| 31 | 🎯 [**Project Page**](https://filmagent.github.io) :octocat: [**Code**](https://github.com/HITsz-TMG/FilmAgent) 📝 [**Paper**](https://arxiv.org/abs/2501.12909) 🧑💻 [**Slides**](https://filmagent.github.io/static/SA24_FilmAgent.pdf) 📽️ [**Video**](https://www.youtube.com/watch?v=hTI-0777iHU) |
| 32 | |
| 33 | </div> |
| 34 | |
| 35 | # |
| 36 | |
| 37 | > [!NOTE] |
| 38 | > This repository also keeps **Video-Claw**, but **FilmAgent is an independent legacy research project** from the lab. It is preserved here for archival and reproducibility purposes only. FilmAgent does not depend on Video-Claw, and Video-Claw does not call FilmAgent. |
| 39 | > |
| 40 | > Current local layout: |
| 41 | > - FilmAgent source code: `FilmAgent/` |
| 42 | > - FilmAgent TTS service: `TTS/` |
| 43 | > - README images: `../FilmAgent-pics/` |
| 44 | |
| 45 | **FilmAgent** is a multi-agent collaborative system for end-to-end film automation in 3D virtual spaces. |
| 46 | FilmAgent simulates key crew roles—directors, screenwriters, actors, and cinematographers, and integrates efficient human workflows within a sandbox environment. |
| 47 | |
| 48 | <div align=center><img src="../FilmAgent-pics/intro.png" height="100%" width="78%"/></div> |
| 49 | |
| 50 | ## 💥 News |
| 51 | |
| 52 | - `2025/2/24`: 🚀 We have integrated **DeepSeek**-v3 and r1 for model selection, allowing for more sophisticated decision-making processes. |
| 53 | - `2025/2/11`: 🎬 We’ve just released a fan-made video about ***NeZha2***, in celebration of its record-breaking success at the box office. Click the image below to watch the video! 👇 |
| 54 | [<div align=center><img src="../FilmAgent-pics/nezha_youtube.png" width="450"></div>](https://www.youtube.com/watch?v=jY3n-AzBtUQ) |
| 55 | - `2025/1/23`: 🙌 We're excited that FilmAgent is recommended by [AK](https://x.com/_akhaliq/status/1882268452716728789), [el.cine](https://x.com/EHuanglu/status/1882294685919772928) and [Theoretically Media](https://www.youtube.com/watch?v=0ebXKegfxWk&t=899s). Thanks! |
| 56 | - `2025/1/22`: 📄 Our paper is now accessible at https://arxiv.org/abs/2501.12909. |
| 57 | |
| 58 | |
| 59 | |
| 60 | ## 🚀 Framework |
| 61 | |
| 62 | Following the traditional film studio workflow, we divide the whole film automation process into three sequential stages: idea development, scriptwriting and cinematography, and apply the **Critique-Correct-Verify**, **Debate-Judge** collaboration strategies. After these stages, each line in the script is specified with the positions of the actors, their actions, their dialogue, and the chosen camera shots. |
| 63 | |
| 64 | <div align=center><img src="../FilmAgent-pics/framework.png" height="100%" width="85%"/></div> |
| 65 | |
| 66 | ## 🌟 Build Your own Film with FilmAgent |
| 67 | |
| 68 | > The following commands assume you are in this `FilmAgent/` directory. |
| 69 | |
| 70 | 1. Install Package |
| 71 | ```Shell |
| 72 | cd FilmAgent |
| 73 | conda create -n filmagent python==3.9.18 |
| 74 | conda activate filmagent |
| 75 | pip install -r env.txt |
| 76 | ``` |
| 77 | |
| 78 | 2. Create `Script` and `Logs` folders under `FilmAgent/`, then replace the absolute pathname `/path/to/FilmAgent` in `main.py` with the absolute path of this `FilmAgent/` source directory and modify the `topic` in `main.py`. Modify the API keys in `LLMCaller.py`. Run the following commands to get the movie script created by the agents collaboratively: |
| 79 | ```bash |
| 80 | cd FilmAgent |
| 81 | conda activate filmagent |
| 82 | python main.py --model "gpt-4o" # openai |
| 83 | # python main.py --model "deepseek-chat" ## deepseek-v3 |
| 84 | # python main.py --model "deepseek-reasoner" ## deepseek-r1 Using this model for a multi-agent process will be very slow, you could try using a single-agent process instead. (Refer to step 6) |
| 85 | ``` |
| 86 | |
| 87 | 3. We use [ChatTTS](https://github.com/2noise/ChatTTS) to provide voice acting for the characters in the script. The TTS-related files are stored in `TTS/`. Download the ChatTTS model files required by `tts_main.py`, then replace `ROOT = "/Path to/TTS"` in `tts_main.py` with the absolute path of `TTS/`. Run the following commands to deploy the text-to-speech service: |
| 88 | ```bash |
| 89 | cd TTS |
| 90 | conda create -n tts python==3.9.18 |
| 91 | conda activate tts |
| 92 | pip install -r env_tts.txt |
| 93 | python tts_main.py |
| 94 | ``` |
| 95 | |
| 96 | 4. Modify the `Script_path`, `actos_path`, `Audio_path` and `url` in `GenerateAudio.py`. Run the following commands to get the audio files: |
| 97 | ```bash |
| 98 | cd FilmAgent |
| 99 | conda activate filmagent |
| 100 | python GenerateAudio.py |
| 101 | ``` |
| 102 | |
| 103 | 5. We now have the `script.json`, `actors_profile.json`, and a series of `.wav` audio files. Next, we need to execute the script in Unity. The recommended version of the Unity editor is **Unity 2022.3.14f1c1**. You need to download the Unity project file we provide [[Dropbox](https://www.dropbox.com/scl/fi/atxhxnqppeofmt471dxr5/TheBigBang.zip?rlkey=o7zuqficetabkk2h1w1npkv3v&st=5s41rzdv&dl=0)][[Baidu Disk](https://pan.baidu.com/s/10wgfvtcG-xnx1fA9Rd_zwg?pwd=5aif)]. After decompression, open `TheBigBang\Assets\TheBigBang\Manyrooms.unity` with Unity. Then replace all the absolute pathnames '/path/to/' with your specific path in `TheBigBang\Assets\Scirpts\StartVideo.cs` and `TheBigBang\Assets\Scirpts\ScriptExecute.cs`. Press **'ctrl+R'** in the unity interface to recompile, click **'Play'** to enter Game mode, then press **'E'** to start executing the script (sometimes the audio files load slowly, so you may need to play it 2 or 3 times before it can run normally). |
| 104 | |
| 105 | > [!IMPORTANT] |
| 106 | > Please add `"com.unity.nuget.newtonsoft-json": "3.0.2"` inside `Packages/manifest.json` if it throws an error, "*the type or namespace name jobject could not be found.*" |
| 107 | |
| 108 | <div align=center><img src="../FilmAgent-pics/unity_1.png" height="100%" width="50%"/><img src="../FilmAgent-pics/unity_2.png" height="100%" width="50%"/></div> |
| 109 | |
| 110 | 6. For the tests on 15 topics in our experimental section, we provide three .py files: `test_full.py` (The full FilmAgent framework, utilizing multi-agent collaboration.), `test_no_interation.py` (A single agent is responsible for planning, scriptwriting, and cinematography, representing our FilmAgent framework without multi-agent collaboration algorithms.) and `test_cot.py` (A single agent generates the chain-of-thought rationale and the complete script). Modify the `model` in these `.py` files, you can try different LLMs. |
| 111 | |
| 112 | ## 🌈 Case Study |
| 113 | |
| 114 | ### 🤝 What does Multi-Agent Collaboration do? |
| 115 | The following table records some comparisons of the scripts and camera settings **before (⬅️) and after (➡️)** multi-agent collaboration, with excerpts from their discussion process. |
| 116 | |
| 117 | <div align=center><img src="../FilmAgent-pics/cases.png" height="100%" width="70%"/></div> |
| 118 | |
| 119 | 📌 **Case Highlights:** |
| 120 | - **Case #1** shows that Director-Screenwriter discussion reduces hallucinations in non-existent actions (e.g., standing suggest), enhances plot coherence, and ensures consistency across scenes. |
| 121 | - **Case #2** shows that Actor-Director-Screenwriter discussion improves the alignment of dialogue with character profiles. |
| 122 | - **Case #3**, in the Debate-Judge method in cinematography, demonstrates the correction of an inappropriate dynamic shot, which is replaced with a medium shot to better convey body language. |
| 123 | - **Case #4** replaces a series of identical static shots with a mix of dynamic and static shots, resulting in a more diverse camera setup. |
| 124 | |
| 125 | ### ⚖️ Comparison with Sora |
| 126 | |
| 127 | <div align=center><img src="../FilmAgent-pics/sora.png" height="100%" width="70%"/></div> |
| 128 | |
| 129 | While Sora (🔗 [Video](https://github.com/user-attachments/assets/65bb4c12-cba0-4ee9-a673-63ea5103fd76)) shows great adaptability to diverse locations, characters and shots, it **struggles with consistency and narrative delivery**, along with **strange artifacts**. |
| 130 | |
| 131 | In contrast, FilmAgent requires pre-built 3D spaces, but it produces **coherent, physics-compliant** videos with strong **storytelling capabilities** (▶️ Watch on [Youtube](https://www.youtube.com/watch?v=yOOycdfolFY)). |
| 132 | |
| 133 | ### ✨ Integration with Text-to-Video Models (e.g. Sora, Vidu) |
| 134 | |
| 135 | We are currently working towards **merging the power of text-to-video models with FilmAgent**, unlocking enhanced storytelling, greater adaptability, and improved consistency. 🚀 |
| 136 | |
| 137 | Here is a preview of our work—a fan-made tribute to the box office phenomenon *NeZha2* [▶️[video link](https://www.youtube.com/watch?v=jY3n-AzBtUQ)]. This video celebrates *NeZha2* surpassing a CNY 8 billion at the box office, officially crowning it as the highest-grossing film in a single territory! 🎉 |
| 138 | |
| 139 | ## 📚 Citation |
| 140 | |
| 141 | If you find FilmAgent useful for your research and applications, please cite using this BibTeX: |
| 142 | ```bibtex |
| 143 | @misc{xu2025filmagent, |
| 144 | title={FilmAgent: A Multi-Agent Framework for End-to-End Film Automation in Virtual 3D Spaces}, |
| 145 | author={Zhenran Xu and Longyue Wang and Jifang Wang and Zhouyi Li and Senbao Shi and Xue Yang and Yiyu Wang and Baotian Hu and Jun Yu and Min Zhang}, |
| 146 | year={2025}, |
| 147 | eprint={2501.12909}, |
| 148 | archivePrefix={arXiv}, |
| 149 | primaryClass={cs.CL}, |
| 150 | url={https://arxiv.org/abs/2501.12909}, |
| 151 | } |
| 152 | ``` |
| 153 |