I am trying to put words to a phenomenon I have been trying to figure out for 2 years. I should have asked for help much sooner.
Is there an adjective for when an output is sharp but so hard for the eyes to focus on? Take for example a person’s hands and feet against a background. Looking at a small part of it, the eyes can discern that this is the person’s hands and fingers only after squinting and crossing eyes for 2 seconds. No part of it is blurry. In fact, it may even be a tad excess sharp. But it takes a whole 2 seconds and the eyes can only focus on that hand area that occupies 1/10th the frame, instead of viewing the hand, connected to an arm, connected to a person, against a background.
All parts of the total frame are distracting to the eyes where it takes 2 seconds to squint on the hand portion, the feet portion, a portion with an object in the background…instead of seeing the frame as a whole.
On a fateful day in late 2024, I overrode my Topaz models configuration from I don’t even remember - a graphics driver update, a corrupted Windows update, etc. The results I had been getting on a 4080 laptop from Stab+Apf2+Prob3+Iris3+Thm2+Ghq5 upscale of 4k to 8k and Stab+Apf2+Iris3+Thm2+Ghq5 upscale of 1080p to 4k were never the same from that day onward. Before, it yielded excellent, clearly discernable outputs. After the incident, “sharp” but muddled, distracting eye pull on exact same settings. I only knew from the Lavf61.3.103 metadata field that I had been between versions 5.0.4 and 5.3.3. I have tried all the following. At this point it feels I have tested everything I have control over, and even wacky extraneous things. Am I missing something?
- Downloading every single Topaz iteration from 3.0 to 5.3.3 models and keeping them when installing new versions (“No” to remove previous models), and trying the retained models on each version between 5.0.4 and 5.3.3 in hundreds of permutations
- Reverting 4080 graphics driver to the version it was in September 2024
- Setting system environment variables to ensure Topaz ffmpeg is being called, and conversely, for it to NOT call Topaz ffmpeg, but rather the packaged ffmpeg 7 that existed in late 2024
- Trying Topaz with various versions of packaged ffmpeg in bin with tens of builds between versions 4 to 8
- Using model manager to download all models in versions between 5.0.4 to 5.3.3
- Changing json for models downloaded to favor fp32 rather than fp16s
- Cycling auto/Nvidia 4080/all GPU options in settings and all memory usage stipulations from 10 to 100%
- Regedit resetting all Topaz settings to default before trying each 5.0.4 to 5.3.3 versions
- Trying other computers. 5090 laptop and 5090 desktop actually produce worse looking distracting outputs than the 4080 laptop
- Factory resetting the 4080 laptop
- Graphics settings making Topaz and ffmpeg exes specifically use, or not use, the 4080 in High Performance mode
- Making presets by entering the exact Proteus and Iris settings from known good metadata in json instead of using GUI sliders
- “Prepping” the AI learning by using a dummy file to run every single setting+model imaginable, in case the AI somehow needs to learn beforehand, despite my process only used Proteus, Iris and Gaia
- Trial/error cranking any json value to 999 that usually have value limits of “1” or “100” for any setting that does not throw an error
