MiniVeo3-Reasoner: Thinking with Videos from Open-Source Priors
## 🎯 Overview
Advanced video models have recently demonstrated remarkable [zero-shot capabilities of visual reasoning](https://video-zero-shot.github.io/), solving tasks like maze, symmetry, and analogy completion through a **chain-of-frames (CoF)** reasoning process.
This project shows that such CoF capability can be **acquired by fine-tuning open-source video models** like [Wan2.2](https://github.com/Wan-Video/Wan2.2).
In the maze domain, the fine-tuned models—dubbed **MiniVeo3-Reasoner**—exhibit **surprisingly strong visual reasoning performance**, achieving **near-perfect accuracy** on in-distribution tests and **robust out-of-distribution generalization**.
Under controlled comparisons, MiniVeo3-Reasoner **significantly outperforms baseline approaches** that reason in other modalities such as text or images.
We further envision that this visual reasoning capability can be **enhanced through reinforcement learning of video models**.
## 🔥 News
- 🚩 **2025.10**: We are thrilled to release MiniVeo3-Reasoner, with mazes as a testbed for visual reasoning!
## 🤗 Models
| Models | Download Links | Description |
| ------------------------- | ------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| MiniVeo3-Reasoner-Maze-5B | 🤗 [HuggingFace](https://huggingface.co/thuml/MiniVeo3-Reasoner-Maze-5B) | Fine-tuned LoRA for [Maze](https://github.com/understanding-search/maze-dataset) tasks (3x3 to 6x6 sizes) from the base model [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) |
## ✨ Examples
Problem Setup
Examples
Maze 3x3
Maze 4x4
Maze 5x5
Maze 6x6
### OOD Generalization
OOD Solution Lengths:
Problem Setup
Examples
Maze 6x6 (solution len > 12)
OOD Maze Sizes:
Problem Setup
Examples
Maze 7x7
Maze 8x8
## 📊 Performance
Following [Visual Planning: Let's Think Only with Images](https://github.com/yix8/VisualPlanning), we report two metrics:
- **Exact Match (EM)** measures whether the model **successfully generates the complete and correct trajectory** that aligns with the shortest optimal valid path.
- **Progress Rate (PR)** measures the **number of consecutively correct steps** (valid forward moves) from the start to the number of steps in the optimal path.
| MiniVeo3-Reasoner-Maze-5B | EM (%) | PR (%) |
| ------------------------------ | ------ | ------ |
| Maze 3x3 | 100 | 100 |
| Maze 4x4 | 100 | 100 |
| Maze 5x5 | 100 | 100 |
| Maze 6x6 | 98.4 | 98.7 |
| Maze 6x6 (OOD solution length) | 53.6 | 59.7 |
| Maze 7x7 (OOD size) | 86.8 | 90.1 |
| Maze 8x8 (OOD size) | 60.4 | 67.8 |
### Comparisons
Under the same amount of training data, we include performance metrics reported in [Visual Planning](https://github.com/yix8/VisualPlanning) for reference and comparison.
| Model | Thinking Modality | Maze EM (%) | Maze PR (%) |
| -------------------------------- | ----------------- | ----------- | ----------- |
| Gemini 2.0 Flash - Direct | Text | 8.3 | 31.4 |
| Gemini 2.0 Flash - CoT | Text | 6.9 | 29.8 |
| Gemini 2.0 Pro (think) | Text | 21.5 | 35.5 |
| Qwen 2.5-VL-Instruct-3B - Direct | Text | 0.5 | 13.6 |
| Qwen 2.5-VL-Instruct-3B - CoT | Text | 0.8 | 8.2 |
| Qwen 2.5-VL-Instruct-3B - SFT | Text | 33.3 | 52.7 |
| LVM-3B - VPFT | Image | 59.0 | 64.0 |
| LVM-3B - VPRL | Image | 74.5 | 77.6 |
| MiniVeo3-Reasoner-Maze-5B | Video | **99.6** | **99.7** |
## 🚀 Get Started
### Environment Setup
```bash
conda create -n miniveo3_reasoner python==3.12
conda activate miniveo3_reasoner
pip install -r requirements.txt
```
We use [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio/tree/ed256ef8be195d5deae2846a7e9f025670d99db3) for diffusion model training and inference. You need also install it:
```bash
git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
git checkout ed256ef8be195d5deae2846a7e9f025670d99db3
pip install -e .
```
### Data Preparation
Our data generator produces a series of mazes with configurable size, path length and amount, outputting a `.mp4` video file and a `.png` image (the first frame of the video).
We use a customized version of [maze-dataset](https://github.com/understanding-search/maze-dataset). You can install it as follows:
```bash
pip install -e data/maze/maze-dataset
```
After installation, use the script below to generate mazes with custom configurations:
```bash
python data/maze/maze_generator.py
```
To reproduce the same data distribution used in our experiments, simply run:
```bash
bash scripts/generate_maze_dataset.sh
```
The result will be in `dataset/maze_train` and `dataset/maze_test` respectively.
### Inference
Download our [LoRA weights](https://huggingface.co/thuml/MiniVeo3-Reasoner-Maze-5B):
```bash
pip install "huggingface_hub[cli]"
huggingface-cli download thuml/MiniVeo3-Reasoner-Maze-5B --local-dir models/thuml/MiniVeo3-Reasoner-Maze-5B
```
To run inference on a single file or directory, use:
```bash
python inference/maze/inference_maze.py [-r] filename_or_directory
```
> 💡 The first run may take additional time to automatically download the base model files.
To perform inference on all test samples, simply run:
```bash
bash scripts/inference_maze_testset.sh
```
### Success Evaluation
Our evaluator compares the predicted trajectory with the ground truth, computing the distance between the two paths.
We implement our own versions of Exact Match (EM) and Progress Rate (PR) metrics for video-based evaluation.
If your generated results are stored in `dataset/maze_test` and named properly, you can evaluate all test samples by running:
```bash
bash scripts/evaluate_maze.sh
```
### Training Models
We train Wan2.2-TI2V-5B with LoRA, following the instructions provided in DiffSynth-Studio. You can easily fine-tune your own models using the same framework.
For your convenience, if you follow ours, you can copy the train dataset `dataset/maze_train` directly into `DiffSynth-Studio/data/example_video_dataset`.
## 🤝 Contributors
[Jialong Wu](https://manchery.github.io/)\*, [Tianhao Huang](https://github.com/MrH2T)\*, [Changjing He](https://github.com/hcjqwq)\*, [Mingsheng Long](https://ise.thss.tsinghua.edu.cn/~mlong/). (\*: Equal Contribution)
We welcome contributions! Feel free to open [GitHub issues](https://github.com/thuml/MiniVeo3-Reasoner/issues) for bug reports or feature requests.
## 💡 Acknowledgements
- [Veo 3](https://video-zero-shot.github.io/): This project is inspired by the impressive zero-shot performance of Veo 3!
- [Wan](https://github.com/Wan-Video/Wan2.2): Powerful open-source video diffusion models used as base models.
- [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio/tree/main/examples/wanvideo): Video diffusion model training.
- [maze-dataset](https://github.com/understanding-search/maze-dataset): Data generation for maze reasoning tasks.
- [Visual Planning](https://github.com/yix8/VisualPlanning): Baseline benchmark for performance comparison.
- [Nano Banana](https://aistudio.google.com/models/gemini-2-5-flash-image): Help in generating the project logo.
## 📜 Citation
There is currently no technical report available.
If you find MiniVeo3-Reasoner useful, we would appreciate it if you could cite our work:
```
@misc{miniveo3reasoner,
title = {MiniVeo3-Reasoner: Thinking with Videos from Open-Source Priors},
author = {Jialong Wu, Tianhao Huang, Changjing He, Mingsheng Long},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/thuml/MiniVeo3-Reasoner}},
}
```
