Nvidia VSR Testing: AI Upscaling and Enhancement for Video
Nvidia Video Super Resolution (Nvidia VSR) was officially rolled out today.beginning Preview at CES 2023, not to be confused with AMD’s VSR (Virtual Super Resolution). The Nvidia VSR aims to do for video what DLSS technology does for games. I agree. Beginners need one of Nvidia’s best graphics cards, namely the RTX 30 or 40 series GPUs. Of course, you also need to set your expectations appropriately.
By now, everyone should be familiar with some of the things that deep learning and AI models can accomplish. Whether it’s text-to-image art generation using things like Stable Diffusion, answering questions and writing articles, ChatGPT, self-driving cars, or many other possibilities, AI is part and parcel of our daily lives. becoming a part.
A basic overview of the algorithm should be familiar to anyone with DLSS knowledge. Get a bunch of paired images. Each pair contains low-resolution and low-bitrate versions of high-definition (and high-quality) video frames, which are run through deep learning training algorithms to ideally upscale and enhance networks teach you how to Converts poor quality input frames to better looking output. Of course, there are many differences between VSR and DLSS.
For one, DLSS gets data such as the current frame, motion vectors, and depth buffer directly from the game engine. Combine this with the previous frame and the trained AI network to generate an upscaled anti-aliased frame. In VSR, there are no pre-computed depth buffers or motion vectors, so everything has to be done purely based on video frames. In theory, VSR could use current and previous frame data, but Nvidia seems to have opted for a pure spatial upscaling approach. But whatever the exact details, let’s talk about what it looks like.
Nvidia provided a sample video showing the before and after output from the VSR.Click here if you want the original 1080p upscale with bilinear sampling source and 4K VSR Upscale Version — It’s hosted in my personal Drive account, so we’ll see how it goes. (If you can’t download the video beyond your bandwidth limit, please let us know by email.)
To avoid potential copyright issues and not include a bunch of our own videos, but to show how it works with other content, some of the output taken from some sports broadcasts I grabbed some screenshots. All we can say is that slow motion videos (such as Nvidia’s sample) give the best results, while fast-paced ones like sports don’t have as much change between frames. It’s more difficult because it can get very large. But in general, VSR works pretty well. Here’s a gallery of some comparative screen captures (captured with Nvidia ShadowPlay).
All these images are 4K JPGs of the highest quality. It’s not lossless, but it can’t be larger than 10MB, so it needed a bit of compression. You can clearly see the difference between normal upscaling (Chrome) and VSR upscaling.isn’t it Large scale It’s an improvement in every sport sample, but it’s definitely sharpening and unblocking, at least in our subjective view.
Clearly, VSR cannot work miracles. Starting with a 720p source he upscales to 4K (9x upscale) is harder than going from 1080p he to 4K (4x upscale). And upscaling from 480p to 4K (20.25x upscaling!) still lacks a ton of detail so you can’t see the strands of the net in either VSR or non-VSR content. Even the Z logo looks much better in the 720p upscaled sample than in the 480p upscaled sample (sorry for the one image overlay).
Good news: If you have an RTX 30 or 40 series graphics card, you can download the latest Nvidia drivers and give VSR a try. A modern Chrome or Edge browser is also required, at least for now. But with the right software VSR seems to work for every video I’ve tried so far.
Bad news: RTX 20 series users are left behind, at least for now. I asked about this requirement, but still don’t have an exact answer regarding omission. Nvidia may have trained the network for Tensor Cores sparsely, which means it currently only runs on Ampere and newer architectures. However, the actual computational load appears to be relatively small, so it seems that Turing compatibility could easily have been chosen from the start if desired.
To demonstrate this, we tested VSR at two different extremes of the VSR spectrum on the same video sequence (a 720p NHL game upscaled to 4K). Above is the RTX 4090 Founders Edition and below is the EVGA RTX 3050. The RTX 3050 is just 73 teraflops, which is also sparse. Both cards actually looked the same. More importantly, we captured power data for the graphics card only.
GPUs | VSR off | VSR on (1) | VSR on (4) |
---|---|---|---|
RTX 4090 (watts) | 28.9 | 32.8 | 36.9 |
RTX 3050 (watts) | 13.0 | 15.9 | 15.9 |
Clearly, neither GPU has been pushed any harder by the VSR algorithm. The 4090 uses 4W more power at VSR quality 1 and 8W more power at VSR 4. In contrast, the RTX 3050 required him 3W more power at either VSR setting. This means that the Tensor Cores are not maxed out on either GPU. This means that even if you have a less mobile RTX 3050 with 4GB VRAM, you can still run VSR.
Overall, this is an interesting take on video enhancement. There are many other algorithms that have been tried to upscale or extend without using machine learning, some may rival his VSR, but the latest his Nvidia drivers and Chrome browser updates Just downloading is not supported. .
As for what it takes to enable VSR, all I have to do is switch to the latest drivers (I received early access to Nvidia’s 531.14 drivers for this review). Under RTX Video Enhancement[超解像度]Check the box and select the desired quality. According to Nvidia, high-quality settings can put more pressure on the GPU, so stick to lower settings if he’s watching a video stream on a second monitor while playing games. is recommended. However, if you’re only watching videos, you can also go completely to yourself and set the quality to 4.