Latte, a novel latent diffusion transformer for video generation, utilizes spatio-temporal tokens extracted from input videos and employs a series of Transformer blocks to model the distribution of videos in the latent space. Latte achieves state-of-the-art performance on four standard video generation datasets FaceForensics, SkyTimelapse, UCF101, and Taichi-HD. paper, code, pretrained
However, Latte still falls short in terms of video generation length and quality compared to Sora. To achieve training and generation effects close to Sora, the Latte model requires more high-quality text-video paired datasets. Therefore, we have created VidFetch, an open-source dataset download tool to obtain copyright-free videos from various free video websites.
| website | windows | macos | linux |
|---|---|---|---|
| Pexels | ✔ | 📆 | 📆 |
| Mazwai | 📆 | 📆 | 📆 |
| Mixkit | ✔ | 📆 | ✔ |
| Pixabay | ✔ | 📆 | 📆 |
| Coverr | 📆 | 📆 | 📆 |
You can install the stable release on PyPI:
$ pip install vidfetchor get the latest version by running:
$ pip install -U https://github.com/heatingma/VidFetch/archive/master.zip # with --user for user install (no root)The following packages are required, and shall be automatically installed by pip:
aiohttp>=3.9.3,
async_timeout>=4.0.3
tqdm>=4.66.2
texttable>=1.7.0
moviepy>=1.0.3
bs4>=0.0.2
selenium>=4.18.1
requests>=2.31.0
texttable>=1.7.0
huggingface_hub>=0.22.2
You only need three lines of code to start downloading the video
from vidfetch.website import MixkitVideoDataset
mixkit = MixkitVideoDataset(root_dir="mixkit")
mixkit.download(platform="windows")