Why fp8 generated video has half size of fp16 ?

[yjhong89]

Hi!

When I run inference script with fp 16 version and fp 8 version, and results are very great.
But, size of generated videos are different for each version and wonder why ??

• Upper one was generated with fp16 and lower one was generated with fp8 model. As seen, fp8 generated one's size is halfed.

[mboboGO]

Hi, thank you for trying out our fp8 version.
As for your question, I have checked the output tensors from both fp16 and fp8 for 256x256x129f video generation. The generated tensors have the same shape and type of "torch.Size([1, 3, 129, 256, 256]) torch.float32".
The generated videos have similar size:

Please give your inference commands and more info, so that I can reproduce your situation.

[mboboGO]

A possible reason is that the fp8 ckpt filters out some high-frequency details that can not observed visually, but the video compress algorithm gives a much smaller file, when saving the video.
This is a just guess ~