Does SeedVR2 Refer to frames beyond the Batch Size?

[overisland]

I have been using SeedVR2 for video restoration via ComfyUI and have achieved excellent results; I consider it to be the best video restoration tool available today!

I have set the batch size to 5 frames.

Previously, when processing long videos, the process would occasionally be interrupted, resulting in a loss of processing time. To mitigate this, I decided to split the original long video into consecutive 5-frame clips and run the restoration process on each clip individually.

However, I observed a obvious degradation in the output quality when applying the process to these short, 5-frame clips, compared to the quality obtained when processing the full, long video.

My initial assumption was that since the batch size was 5, splitting the video into 5-frame segments would produce the same result.

Given the observed quality degradation, I now suspect that SeedVR2 references frames beyond the set batch size (5 in my case) to get context when a long video is provided. Is this correct?

Thank you!

[adrientoupet]

Hi there - if this is related to the ComfyUI integration, you might want to post that on the other repo https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler

But to answer your question, check your SeedVR2 settings, are you using temporal overlap? That could explain why it looks better when processed together even if the batch is similar. Either way, share the full debug log on the other repo and we'll help you out.

[overisland]

@adrientoupet Thank you for your advice. I will do so.

Btw I'm not using temporal overlap.