I just had a look and Proteus 1080p to 4K only uses about 4GB of VRAM. Therefore, even 8GB of VRAM should be plenty for your use case.
I also suspect that with your current CPU/RAM, a powerful GPU might not be utilized fully. U can use GPU-Z Graphics Card GPU Information Utility to check on your current GPU utilization.
Maybe u want to consider to buy a used GPU, sth like a 3080 at 300 euros is about 20% faster than a 5060Ti but 30% cheaper on ebay than a new 5060Ti. A 3060 might cost only about 200 euros on ebay but but a 3080 should be more than 1.5 times faster (if your CPU/RAM can make use of it).
Sorry, start again.
I am looking to upgrade my graphics card, the card I have is great for Iris & Protues, about 20-25fps, but I am talking about the new ‘Starlight Mini (Local) AI Model’ that has come out, that I thought this thread was about.
So, for the Local Starlight Mini Model, upscaling to a max of 1920x1080, is it the memory or the speed of the card that is the most important factor.
Is my money better spent on 16Gb of RAM or the fastest card my budget can buy with 8Gb.
I haven’t tested the the official TVAI 7 yet, but in the beta, some videos could be processed at the original resolution and the speed was as fast as it can go. For me that was 0.7 to 0.9 fps. If they have fixed the randomness of the output resolution, then you could potentially run a Starlight Mini pass and then something like Gaia or Proteus to upscale it. The overall time taken should be less than pure Starlight Mini to upscale. Of course the results might not be the same. We’d have to test it out to see if they’re better or worse.
Btw, there is also something weird going on with the available output resolutions on portrait rotated videos. As an example, the smallest possible upscale size for a 360 x 640 sized portrait video is 1280 x 2276 (Minimum).
I understand that, but why is it then that a landscape orientated video with the same resolution (640 x 360) has a minimum upscale size to 1280 x 720? Shouldn’t it then be possible to upscale a portrait orientated video to 720 x 1280?
I import 720x576 25fps mp4 file. Export it with Starlight Mini x3 resolution, with the same 25fps and my output h264 file is 9 frames SHORTER than the original! 8 frames at the back and 1 in front. Any ideas?
This is awesome, thanks. I suggested this type of process months ago. It is the logical way to upscale videos properly.
You said “the larger the video, the more the model will know”. Based on your testing, does that mean that exporting 5 seconds of a video will always be worse quality than exporting the entire video because it only analyzes those 5 seconds of frame data that I am exporting?
Or…if I export 5 seconds of the video, will it still look at frame data from the entire video, including frames that I’m not exporting?
Hello,
Regarding Starlight Mini, when you say in GPU requirements that: NVIDIA GPU with at least 8GB of video memory, but we strongly recommend a GPU with 16GB or more memory for optimal performance and quality, it is understandable for performance, but as a person looking for optimal quality and a holder of RTX 3070 8GB VRAM, how is the result quality impacted? Compared to performances, quality is a critical point for me to consider upgrading my hardware.
Thanks.
FYI, you can find patents by search on Google patents (https://patents.google.com) or at the USPTO. Also, US patent applications are published 18 months after the earliest filing date (or earlier at applicant’s request), so you don’t have to wait for a patent to be issued to inspect a filing.
