Saint-Archives.com Forum

Thanks, Adam

Yes, as I said, the smaller photos aren't great, so need a bit more work.  They were lazily fed through Topaz with the same settings as the larger photos, but actually need more individual attention.

With 077 as the example, I think somewhere in between my version and yours would be a good place. 

Yours has maybe gone a little to far on blurring out the detail, whereas mine has brought out too much detail. 

They could also do with meeting in the middle in terms of colour balance.  Mine's a bit yellow-biased wheras yours is red-biased.  

077 was tricky because it came from the LegActionMag series which was darker than the ViewPornstars series, so we have both tried to adjust the colour balance and contrast and come up with different results.

I'll have another go at the smaller photos individually.

I won't have time to play with photoshop today, but you're right about my picture : too red. I could simply low saturation balance a little :

That's much better but i don't like at all your 082 and 086 : 086 because Silvia's face is too white and looks cadaveric, and 082 because there is a grey pattern on her body.

Perhaps you could work again on these 2 or play with my upscales :

My next attempt at upscaling the LegAction.com photos (077-092)

I realised I had gone down a cul-de-sac with the colour balance of my previous efforts, so I went back to the originals and using a different method to try to get the colour balance more in keeping with the ViewPornstars set.  I'm a lot happier with the results.

set088_077	set088_078	set088_079	set088_080	set088_081
set088_082	set088_083	set088_084	set088_085	set088_086
set088_087	set088_088	set088_089	set088_090	set088_091
set088_092

 P.S. I have updated the original post with these versions.

To be honest, Topaz can do better with the facial detail, but it can also overdo it, making Silvia look strange.

I deliberately set the facial reconstruction at a low value (25%) to keep Silvia's face recognisable, but perhaps I aimed too low.

Here are some tests from 077 to show what I mean.  The percentages are the levels of facial reconstruction set in Topaz.

0%	50%	100%

All look pretty much identical in thumbnail form, but you can see obvious differences when you view them at full size.

0% lacks detail as you would expect. 

50% is better. Silvia looks natural, but the detail still slightly fuzzy around the eyes. 

100% has sharper detail around the eyes, but in my view has over-emphasised Silvia's teeth, making her look a little buck-toothed. 

Both 50% and 100% versions have also blended out much of the tonal differences in Silvia's skin.  It's that effect I was mainly trying to avoid, as it looks out of place, like a faked photo, which is why I set the target so low.

However, we can use the different levels of facial reconstruction selectively by overlaying the images and adjusting their alpha levels to blend them together. 

Here is a composite of all three images done in this way.

The bulk of Silvia's skin is from the 0% image to keep the tonal range, her teeth from the 50% image to reduce the emphasis, and her hair, eyes and nose from the 100% image for maximum detail. These are all fairly subtle differences, but I think the result is an improvement on all three individual photos.

I'll have another go using this composite method, if you think it's worth it. 

Or please tell me if I'm barking up the wrong tree... 

Thanks, Adam.

I suppose the AI models are in their infancy. Human facial recognition is very highly tuned, so tiny facial imperfections are magnified in our perception, whereas similar differences in other areas of the photos are almost imperceptible.

There are only a few small images in this set that require this manual treatment.  I've come this far with it, so I think I'll persist with composites.

Where's the fun if it's easy?