| --- |
| license: mit |
| --- |
| |
|  |
|  |
|
|
|
|
| # WorldModel |
| Note: This was a passion project, crafted in two days. I hope it can be of help to many! |
|
|
| WorldModel is a MaskGIT model trained on 8x8x8 Minecraft voxel volumes. Beyond generating blocks from scratch, |
| it excels in filling spaces based on neighboring blocks, ensuring seamless integration in Minecraft worlds |
|
|
| The model weights can be found [here](https://drive.google.com/file/d/1--4Z5VQ-mRQz805yDmJPsPKrIrVCaoCv/view?usp=sharing) or in this repo |
|
|
| ## Prerequisites: |
| Ensure your server has the mcpi package installed. This code relies on it to place blocks. |
|
|
| ## Functionality: |
| This tool uses a special mask token, designated as 831. When you provide a tensor of shape [1,512] containing |
| block values ranging from 0-831(including the mask token), the generate function will intelligently fill in |
| the masked areas until none remain. |
|
|
|
|
| ## Sampling Strategy Adjustments to maskGIT: |
| In contrast to the original maskGIT, sampling solely from the most confident tokens can lead to mode collapse, often |
| resulting in generating only stones or air. To address this, I've introduced a random sampling strategy. In the early |
| sampling phases, the confidence of potential tokens is disregarded, and tokens are selected for unmasking at random. |
|
|
| You can fine-tune this behavior using the "random_steps" variable. It's crucial to strike a balance between "random_steps" |
| and other parameters such as the total "timesteps", "temperature", and "topk_filter_thres". |
