I discovered while running the Precise model is that on my 5090 it’s using an Ampere series SM80 and BF16. There could be an option for FP8 and even FP4 for the 5090, which would greatly speed things up. I couldn’t see it for the other models because they were disabled after I reinstalled.
And degrade Quality.
The only way to know the cost-benefit ratio is to experience losses during testing. That’s why I asked for options to be added for the user. Those who don’t want to test can simply not choose the option.
Looking at the quality of e.g. SeedVR2 with FP8 models I don’t even want to test this..
Higher precision models do deliver quite a bit higher quality output (eg 3b Q8 really is better than the FP8 model).
The only way nvidia did speed up ai from 2016 until now was mainly by lowering quality.
To the point where we did accept that lower quality is ok.
Now we talk about faster and faster but for what tradeoff.
Lets see what will happen.
Take a look at these studies on Blackwell’s new MXFP8 and NVFP4 technologies. It’s quite interesting.
Yes but its all related to Large Language Models and not Image Generation.
An the Medium article, i did watch some of his videos.
When he compares his software to others he don’t speak about wich models have been used.
SECourses Upscaler Pro Beating Topaz AI by Far With Specalized FlashVSR+ & SeedVR2.5 - Local Windows
