Nuke 15.1 | Machine Learning Updates

Welcome to the latest update of CopyCat in Nuke 15.1, featuring improvements to performance and usability.
The CopyCat Runs table now includes a new training menu, making it easier to manage multiple training runs. The new menu adds the option to quickly toggle the visibility of, delete, create inference from, and resume training checkpoints.
We’ve also introduced mixed-precision training, boosting training performance by using half-precision floating-point operations for non-critical layers while maintaining full precision for critical ones. This optimizes GPU memory usage and makes training up to 2x faster without degrading output fidelity.
Experience a more functional and efficient CopyCat with Nuke 15.1, speeding up your training process without sacrificing quality.
Download Nuke 15.1: campaigns.foun...
Learn More: learn.foundry....
Questions or Feedback? Leave a comment below or join our community community.foun...
About Us:
Foundry creates award-winning creative software so global studios can turn incredible ideas into reality.
Subscribe to our channel and get the latest news, tutorials, webinars and updates from the Foundry team.

[@Arminkenedi]

It's only worth it if a shot is long. This means most of the time, the Rotobrush from After Effects is much faster bcz the shots are less then 2 sec. Copycat takes too long to implement! Perfect rotoscoping up to 10 frames! In the end, you spend more time than if you manually rotoscoped everything. I don't know whose idea CopyCat was, but it's really pointless. I have so many arguments about why CopyCat is so bad.

[@ziadramadan7819]

صباح الفل

[@panosdimi4713]

Thank you ❤

[@alptonay1481]

I tested the new copycat last friday. I must say it's definetely improved! Is it faster? An absolute YES! Is it more accurate or precise? I was impressed! By the way, I tested the copycat for the beauty work and de-aging tasks we already did before Nukex V14. When I applied the *.cat file to the batch shots consistency was amazing. When extra taraining needed for the batch shots using the previous training as a checkpoint gave the advantage in a better way and just 2000-3000 epochs were enough to reach the result. Roto is not my concern, so I can't really say anything about that part.

[@Arminkenedi]

R u wokring with Nuke. Bcz it sound like another Software. Iam using the last version too. Its useless. Too long too hard to make it good and tooo hard for a 2 sec clip!

[@alptonay1481]

@@Arminkenedi Hello. yes we are working with NukeX. But I must mention again I worked with batch shots so it contains multiple shots. And we are using ML tools for complicated shots which also need consistency. So 2 sec clip is not the case here unless it's a very very complicated shot.

[@Arminkenedi]

​@@alptonay1481 That is exactly what I am talking about! How many shots are long and very complicated? Very few, and most of the time-actually, in 90% of the cases-a clip is under 5 seconds und they are not super complicated. Is the point of CopyCat a function for very few shots?