Photo AI 1.0.7 - is it just me, or are the results vastly improved?

Zeee wrote:

Digital Nigel wrote:

Zeee wrote:

Digital Nigel wrote:

Docno wrote:

Digital Nigel wrote:

Zeee wrote:

Docno wrote:

Digital Nigel wrote:

logatom wrote:

Hi,

Great output from DxO, but obviously it's not coming from the RAW files you linked to. Would you please share the correct file, if possible?

Sorry for the incorrect link — this is the correct file:

https://www.dropbox.com/s/sm5eqr025xs9bei/RX604368%20Temple%20of%20Saint%20Sava%2C%20Belgrade.ARW?dl=0

This is what I get. I processed the Raw in the standalone and noticed, with default settings, the output was about the same as what you got. Then I tried it using the Photoshop plugin. Interestingly, Photo AI defaulted to 'Strong' noise reduction, which I think is what was causing the result you got. I switched this to standard (light, or whatever the weaker version is), and I got a much better outcome. I then did a simple (auto) contrast/brighten. What do you think?

This is what I got with Photo AI opened from LrC as a plug-in and edited as RAW.

I've compared these two attempts with PL6 and my original PAI stand-alone:

Four-way comparison: Docno's version is soft and lacking in detail, Zeee's version is over-sharpened and full of ugly artefacts (look at the saints' faces). The three Photo AI versions can't even agree on colour.

Even with the help of PS and/or LR, Photo AI still can't produce the same quality as PL6 does alone, quickly and easily. There's less detail and/or more artefacts in all the Photo AI versions.

Except you’ve zoomed in more on my version which invalidates the comparison. I’m not saying you’re wrong, but you can’t compare image detail at different zoom factors

I didn't zoom in more: your upload was up-sized (why?). All four examples are compared full-size as uploaded.

But the result would be the same even if you hadn't up-sized it: there simply is less detail in the Photo AI results. Zeee tried to compensate by sharpening it, but that added ugly artefacts, not detail.

I wasn't trying to compensate. I just wanted to see if that mush was it. I normally do not apply sharpening in in LrC after doing so in Topaz do I normally export at screen high. I put little effort into it. This just showed me I can do better than what you posted.

I posted the output of stand-alone TPAI, to show what it looked like (people had been claiming that it was a good tool for developing raw files). You wouldn't get anything very different from TPAI.

So you used LR to improve the image, and anyone using TPAI would similarly need to use some sort of photo editor to do all the many things TPAI doesn't do. They can certainly improve the lighting, contrast, colours, levels, etc. They shouldn't need to do any sharpening, given that TPAI includes what is supposedly top quality AI-based sharpening.

But none of those photo editors, however good, can restore the detail that TPAI lost. Over-sharpening in LR as you did might make the image look sharp at first glance, but just adds artefacts, not detail. It's impossible for LR (or PS, AP, PSE, PSP) to replace what TPAI missed.

Did you apply lens corrections with PL because that is like applying sharpening with other software.

Of course, as you can see from the image. I always use PL's excellent lens corrections.

Also DeepPrime just does not remove noise. I had PL4 which I can't use but I could see that it improved the image.

Yes, and DeepPRIME XD seems to improve it further. The results are even cleaner, sharper and more detailed.

So based on that who made the rules that I shouldn't have to use other software to finish.

I agree that other software is needed to finish the processing, either before and/or after Photo AI.  But use of that software couldn't correct for the loss of detail in TPAI (or 'mush', as you accurately describe it).

It designed to work as a plug-in for LrC, PS and other developers so why can't I use the tools in the software. I've read quite a few posts of people using Affinity to finish files from PL. Why do they need it?

Because TPAI only does much less than half the work of processing an image (just NR and sharpening). So of course you need some other photo editor to complete the task. But that can't make up for TPAI's deficiencies as a raw developer (poor lens correction, cropped image, severe loss of detail, soft).

You also said at first glance it looked OK so it wasn't glaring, which was my whole point. If I posted that full size file I really doubt someone would notice or care if those saints looked a little rough.

It's still twice the effort, for a much worse result. And they'd certainly notice the unnecessary crop in your full image.

I did say I pushed it harder because I just wanted to see and I'm not worried about it taking longer than PL. I've got time. I posted ON1 will be coming out with Tack Sharp and I got heck from someone telling me we are relying too much on AI. I'd put DeepPrime in the AI category, not that there is anything wrong with that.

DxO itself says it uses AI, based on its collection of millions of test images. Presumably, during its camera and lens tests, DxO shoots the same subjects at varying ISO levels, and so has built up a huge database of clean, low ISO and noisy, high ISO raw images of the exact same scene. The machine learning can use these millions of cases, and detailed test data about that particular camera sensor and lens, to predict exactly what the clean, detailed, sharp version of any fuzzy, noisy raw image should look like. It certainly works like magic.