| # Visualization | |
| This repository contains visualization tools for the LLaDA project. | |
| ## Implementation Steps | |
| ### Step 1: Generate Sampling Process | |
| Run `generate.py` to produce your own sampling process records. A sample output (`sample_process.txt`) is included for reference. You have the option to: | |
| - Utilize the provided generate.py script | |
| - Modify both the prompt and generation parameters | |
| ### Step 2: Generate Visualization HTML | |
| Choose between two visualization styles: | |
| - **Paper Style**: | |
| `visualization_paper.py` produces visualizations matching the format in [our arXiv paper](https://arxiv.org/abs/2502.09992) | |
| - **Zhihu Style**: | |
| `visualization_zhihu.py` generates visualizations compatible with [Zhihu's format](https://zhuanlan.zhihu.com/p/24214732238) | |
| The scripts will: | |
| 1. Automatically create an `html/` directory | |
| 2. Generate individual HTML files for each sampling step | |
| *Note: The current implementation defaults to 64 sampling steps. | |
| ### Step 3: Create PNG Sequences | |
| Convert generated HTML files to PNG format for GIF creation. These image sequences can be used with any standard GIF generator to visualize the complete sampling process. | |
| ## Technical Notes | |
| - Ensure Python 3.8+ environment | |
| - Install required dependencies: `pip install html2image` | |
| - For custom configurations, modify constants at the beginning of each script | |