This weekend we rebuilt our development environment from the ground up. This includes updating graphics driver, CUDA, cuDNN and recompiling multiple layers of software to now support the latest line of graphics cards from NVIDIA, the RTX30xx series (RTX3070, RTX3080 and RTX3090). In more technical terms it means we are now supporting compute capabilities 8.0 and 8.6 (please see https://en.wikipedia.org/wiki/CUDA). The newly built version variants of NNSuperResolution are available from the Downloads page as usual. If you do indeed own a RTX30xx card, be sure to look for the versions compiled against CUDA11.2.
In the middle of January we did a series of before & after examples of the output of NNSuperResolution on Instagram and Facebook. For easy access, we’ve also posted these examples on this page here on the website.
We have created a dedicated page to make it easy to request a free trial license for NNSuperResolution. You can request either node locked or floating licences. The default is that we create a free test license for you that will expire after 10 days. This way you can test the plugin fully, without any watermarking/noise, to properly be able to evaluate the results for yourself on your own material.
After talking to some clients about their installation experience, we’ve decided to also provide downloads of the NNSuperResolution plugin bundled with the needed CUDA and cuDNN libraries. This will make for a much easier installation procedure if you don’t previously have the NVIDIA CUDA Toolkit and NVIDIA cuDNN libraries installed on your system. The bundled CUDA & cuDNN libs can be installed into the same NUKE_PATH directory as the main “NNSuperResolution.so” is installed into, and the plugin will find and use them directly from there. These new versions are available on the downloads page.
We are continuing our development journey towards finding a good super resolution solution for sequences that will produce a much more temporally stable result. While the translation invariance loss, from the previously mentioned paper, does help with producing a more stable result in general it doesn’t produce as temporally stable sequences as we want. We are currently looking into the methods presented in the paper “Learning Temporal Coherence via Self-Supervision for GAN-based Video Generation“.
It’s been a week since our release of NNSuperResolution, and we are happy to have received a lot of encouraging comments from artists all across the globe, especially in this thread on LinkedIn. The next thing on our agenda is to look into how we can potentially make it more temporally stable. As mentioned on the product page, the current super resolution solution is based on a still frame trained neural network. This is resulting in very nice and sharp high-res still images, but doesn’t necessary result in smooth images sequences / video. Image detail may stutter and flicker between frames, all depending on how the input images are looking. There are several different approaches available in recent research to try and achieve a more temporally stable result without degrading the upscaling results too much. We are currently looking into the methods presented in the paper “Single-frame Regularization for Temporally Stable CNNs“.