This is currently the best open‑source video model, yet the conservative algorithm of the VAE seems to create a computational bottleneck?
Thank you, great Chinese friends! This is currently the best open‑source video model, yet the conservative algorithm of the VAE seems to create a computational bottleneck. Our tests suggest that the VAE model has a hard‑coded highly conservative 256‑chunk setting and does not accept external computational chunking. As a result, encoding and decoding can only run at a very low speed even on high‑performance GPUs, leading to poor GPU utilization efficiency. Of course, all the above are merely observations and speculations from our tests.
We hope the official team can release optimized VAE algorithms and models if possible. Thank you very much! You guys are amazing!
感谢你们，优秀的中国朋友们！这是目前效果最好的开源视频模型，但VAE保守的算法似乎造成了计算瓶颈。我们的测试发现，VAE模型内置了一套高度保守的256分块设置，并且不支持外部计算分块。因此，即便使用高性能GPU，编解码也只能以极低的速度运行，造成GPU利用率低下。当然，以上内容都仅来自我们测试得到的观察与推测。
我们希望官方团队如果可以的话，能够发布经过优化的VAE算法与模型。非常感谢！你们做得太棒了！