Model update and resolution

[yjr0229-art]

Hi, thank you for your great work! I'd like to know when will the DVD1.1 release? And I also wonder does DVD, like DAV2, have good scaling capability at higher resolutions than training during inference?

[soyouthinkyoucantell]

Hi @yjr0229-art ,

Thanks for your kind words!

The DVD v1.1 is about to be released within one month. You could expect better quality and robustness against reflective surfaces.

For the latter, although our method can produce finer details at higher resolutions, we do not recommend it, as the computational cost of the attention module increases quadratically with spatial resolution. However, we will add an SR module with the release of DVD v1.1, so you can expect finer detail at a negligible increase in computational cost.

Hope this answers your questions.

Best,
Faye

[QuickscopingFTW]

i see you mentioned a SR module, im guessing you are referring to a super resolution step, out of curiosity does this step run at the same time as each segment of the video is processed or after each segment, or does it do the upscale step after the whole video is processed? i was wondering because at the moment on of the bigger issues with the generation is the shift in scale between the around 81 frame processing windows, if the upscale step was processed right at the end after everything has fully generated could the upscale step also act as a kind of smoothing feature that smoothed out the transitions between the segments? would love to know if this is at all possible, thanks in advance