The new 256 bit model

[JosTheBossX]

Hi,

I just saw that there is a new model being uploaded here i.e. the 256 bit one. But I think in the paper only the 96 bit one is mentioned. I know that this was a new update this year in March, but do we have any results displayed somewhere on the bit acc and the quality metrics of how well it performs w.r.t its 96 bit version?

I felt that this is really interesting having this big of 256 bit decoder.

[pierrefdz]

Hi,
There is no paper yet.

Results with regards to the bit accuracy should be very similar to the 96bits version, maybe a bit lower for the combined augmentations with video codec augmentations + crop and brightness, since the current 256b model is trained on images only (but you can increase a bit the strength of the watermark distortion to improve robustness to these augmentations).

You can find the detailed metrics.csv -- obtained on high resolution videos -- at the end of: https://github.com/facebookresearch/videoseal?tab=readme-ov-file#full-evaluation

[JosTheBossX]

Thanks! I will have a look at it.