Quote:
Originally Posted by Bassnut
Excellent image Andy!.
Based on forum usage reports, I tried Topaz gigapixel AI, it was apparently better than Denoise AI and Sharpen AI combined for astropics. To my surprise, binning X2 an image and then gigapixel up scaleing it back to native res gave stunning results !!. (the algorithm denoises and sharpens too). Geez you know, is this fake, if it actually reasonably "sharpens" blurred bits in a sort of deconvolution fashion, then this has some scientific credibility, convolution is a mathimatically reversable effect. That binning helped (halving the res) is interesting, perhaps this allows more accurate processing with reduced noise.
|
Hi Fred there is a massive difference between an algorithm using maths like a PSF for deconvolution for example and a neural network trained to do a task by using a huge pool of data points. AI has its place. If Google wants to develop a self driving car that "knows" what a kerb, a road, a tree or another car is then it has a huge amount of images with street view or even all the pics people post online to train a machine to recognize potential obstacles. Then you call it machine learning and brand the end product AI. I suspect the star removal "AI" was initially fed a huge amount of starfields for a given focal length so it "knows" what a star is in order to remove it. Again it is trained and only as good as the data it was trained with. As far as a Sharpen AI goes you can't make sh!t up out of nothing and calll it a day. So the more you train it, the better the results will be? It works 100% giving you the exact same result for the same settings every single time, otherwise it's unreliable.