I have a ton of older videos in various forms, Hi8. Digital8, MKV, MP4 etc etc. I have tried every way I can think to improve taken screenshots and stills from these videos with Photo AI but the software has little to no affect on it. Can someone explain like im 5 why this is?
Before anyone suggests it my PC simply doesnt have the computing power to use videoAI to improve the sources and I dont need better quality videos, just still from them. So what is inherently different about a screenshot from a video saved as a JPG as opposed to a photo JPG and does anyone have any suggestions of methods or alternative software to make screenshots look better??
Maybe make the pixel size of the screenshots smaller to bring the detail density closer to the pixel density. This helps the a. i. to work on the material. For example: make a 1920 x 1080 image smaller like 1280 x 720 or 768 x 462.
The performance of Photo AI depends on the quality of the original image. The AI models work based on existing image data so if the quality of the original is poor or lacks detail, then the app will struggle as well.
It sounds like that’s a potential cause because you mentioned that these videos are older. If you want, you can upload one of the screenshots and I can give some workflow suggestions!
Screenshots from videos are usually heavily compressed , oversharpend and with motion blur (due to slow shutter speed.)
Video compression artifact are not the same as JPG compression artifact because the way of compression are not the same. Photo AI don’t have a specific AI model for screenshots. And their standard model are not trained with screenshots photos.
If the compression artifact cannot be removed by PAI, you may try to reduce the resolution of the screenshot image. For example, 1920x1080 → 960x540
Because Video marcroblock size can be 64x64, 32x32, 16x16, etc. while JPG macroblock size is 8x8. Reduce the resolution make the video block artifact “similar” to JPG block artifact , hence easiler for PAI to remove.
Video AI
VAI support image sequence input (min. 4 images) and export as image sequence.
At first, I have to state that I don’t have Photo AI but Gigapixel AI which I use on video stills. Since I assume the algorithms of PAI and GPAI may be similar, here are my perceptions on the matter.
I have a Sony AX 53 video camera which produces MP4 XAVC Video. While the camera is capable of 4K, I, for several reasons, usually record Full HD (1920x1080).
I sometimes feed stills of the videos into GPAI, enlarging them to about 9 Megapixels, and it is pretty amazing how good the results are in many cases. If the results are below my expectations, it is usually due to motion blur or increased noise in subpar light conditions. In fair weather conditions, neither of the latter affect the stills in a way that makes them unsuitable fĂĽr GPAI. But, as it was written before, especially motion blur can be a problem with stills taken from video recordings.
I have also tried to feed stills from 1280x720 TV recordings and even Youtube videos into GPAI but the results do not keep up with those from videos taken with the AX 53. The high compression ratios used by many Youtube videos seem to be a show stopper as well as the high amount of motion blur found in many TV recordings. So, yes, GPAI has its limitations as to processing video stills (and so does PAI, I guess) but it can be done provided that high quality video is fed into it. As others mentioned, many of your older videos may just contain too much motion or compression artefacts. I have tried 720x576 interlaced from SD satellite TV recordings or from digitized VHS tapes, and neither were suitable for being improved by GPAI. The ratio of textures appearing in those recordings to noise/compression/motion artefacts just seems too bad to get a decent result.