Topaz Photo AI

37 views
Skip to first unread message

Alfred Molon

unread,
Sep 22, 2022, 3:10:48 PMSep 22
to
Some tests with Photo AI (images shot with the Olympus E-M1 Mark II and
III cameras):

1. The kingfisher I shot in 2019 with the Olympus 75-300 lens at 300mm /
ISO 6400.
Left is my RAW processing result of 2019, right the same RAW just
re-processed with Photo AI:
https://www.molon.de/images/topaz_vs_dxo/kingfisher_photo_ai.png

2. A red leaf monkey I shot last month in the Danum valley (ISO 6400,
Olympus 100-400 at 400mm).
Left is the out-of-camera JPEG, right the output of Photo AI:
https://www.molon.de/images/topaz_vs_dxo/red_leaf_monkey_photo_ai.png

Quite impressive detail recovery, especially in the monkey picture.

The 75-300 is a relatively inexpensive tele lens, and ISO 6400 is not
exactly an ISO level you should use with a micro 4/3 camera. But with
modern software, no problem, you get a sharp and noiseless result.
--
Alfred Molon

Olympus 4/3 and micro 4/3 cameras forum at
https://groups.io/g/myolympus
https://myolympus.org/ photo sharing site

HunterBD

unread,
Sep 22, 2022, 3:58:06 PMSep 22
to
On 22/09/2022 20:10, Alfred Molon wrote:
> Some tests with Photo AI (images shot with the Olympus E-M1 Mark II and
> III cameras):
>
> 1. The kingfisher I shot in 2019 with the Olympus 75-300 lens at 300mm /
> ISO 6400.
> Left is my RAW processing result of 2019, right the same RAW just
> re-processed with Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/kingfisher_photo_ai.png
>
> 2. A red leaf monkey I shot last month in the Danum valley (ISO 6400,
> Olympus 100-400 at 400mm).
> Left is the out-of-camera JPEG, right the output of Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/red_leaf_monkey_photo_ai.png
>
> Quite impressive detail recovery, especially in the monkey picture.
>
> The 75-300 is a relatively inexpensive tele lens, and ISO 6400 is not
> exactly an ISO level you should use with a micro 4/3 camera. But with
> modern software, no problem, you get a sharp and noiseless result.


Thanks, Alfred. VERY impressive. 🙂

I've just had a look here:-

https://www.topazlabs.com/topaz-photo-ai

sobriquet

unread,
Sep 22, 2022, 7:00:15 PMSep 22
to
With AI these days you don't even need to take pictures.. AI can just mash up custom images from
images found online.

https://thispersondoesnotexist.com/

Bill W

unread,
Sep 22, 2022, 7:52:08 PMSep 22
to
On Sep 22, 2022, Alfred Molon wrote
(in article <Qq2XK.255858$YC96....@fx12.ams1>):

> Some tests with Photo AI (images shot with the Olympus E-M1 Mark II and
> III cameras):
>
> 1. The kingfisher I shot in 2019 with the Olympus 75-300 lens at 300mm /
> ISO 6400.
> Left is my RAW processing result of 2019, right the same RAW just
> re-processed with Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/kingfisher_photo_ai.png
>
> 2. A red leaf monkey I shot last month in the Danum valley (ISO 6400,
> Olympus 100-400 at 400mm).
> Left is the out-of-camera JPEG, right the output of Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/red_leaf_monkey_photo_ai.png
>
> Quite impressive detail recovery, especially in the monkey picture.
>
> The 75-300 is a relatively inexpensive tele lens, and ISO 6400 is not
> exactly an ISO level you should use with a micro 4/3 camera. But with
> modern software, no problem, you get a sharp and noiseless result.

I already had the first 3, and they gave Photo AI to everyone who had them
already. The good results can be crazy good, and the bad results are just,
“oh well”. It’s AI, so anything can happen. It’s just another tool
but I think it’s a pretty good one overall. My only worry is the company
itself. I have no idea how to get any support if I need it, and no idea if
they’ll still be in business tomorrow.

RJH

unread,
Sep 23, 2022, 12:52:44 AMSep 23
to
On 23 Sep 2022 at 00:52:00 BST, Bill W wrote:

>> But with
>> modern software, no problem, you get a sharp and noiseless result.
>

Agreed, impressive.

> I already had the first 3, and they gave Photo AI to everyone who had them
> already. The good results can be crazy good, and the bad results are just,
> “oh well”. It’s AI, so anything can happen. It’s just another tool
> but I think it’s a pretty good one overall. My only worry is the company
> itself. I have no idea how to get any support if I need it, and no idea if
> they’ll still be in business tomorrow.

It's the 'free updates for a year' bit, as well as the high (for me) intitial
price. Effectively makes it a subsciption model.

Is there something similar as freeware or relatively inexpensive?

--
Cheers, Rob, Sheffield UK

David Taylor

unread,
Sep 23, 2022, 1:53:24 AMSep 23
to
On 22/09/2022 20:10, Alfred Molon wrote:
> Some tests with Photo AI (images shot with the Olympus E-M1 Mark II and
> III cameras):
>
> 1. The kingfisher I shot in 2019 with the Olympus 75-300 lens at 300mm /
> ISO 6400.
> Left is my RAW processing result of 2019, right the same RAW just
> re-processed with Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/kingfisher_photo_ai.png
>
> 2. A red leaf monkey I shot last month in the Danum valley (ISO 6400,
> Olympus 100-400 at 400mm).
> Left is the out-of-camera JPEG, right the output of Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/red_leaf_monkey_photo_ai.png
>
> Quite impressive detail recovery, especially in the monkey picture.
>
> The 75-300 is a relatively inexpensive tele lens, and ISO 6400 is not
> exactly an ISO level you should use with a micro 4/3 camera. But with
> modern software, no problem, you get a sharp and noiseless result.

Thanks for posting! Impressive, yes!

With the second (Monkey) image, was Topaz processing the OOC RAW data, or the
JPEG image shown on the left?

The Web page seems to give little information on the product - will it process
Panasonic RAW format as well as Olympus format?

Sadly my graphics card is just an NVIDIA GeForce GT 710 with 2 GB, so it's not
supported. A pity, but at least it saves me US $199/159!

--
Thanks,
David
Web: https://www.satsignal.eu


HunterBD

unread,
Sep 23, 2022, 2:54:43 AMSep 23
to
Fascinating! :-D

Did you explore this (pop-up)?

Imagined by a GAN (generative adversarial network)
StyleGAN2 (Dec 2019) - Karras et al. and Nvidia
Don't panic. Learn how it works [1] [2] [3]
Code for training your own [original] [simple] [light]
Art • Cats • Horses • Chemicals • Contact me
Another | Sponsor

HunterBD

unread,
Sep 23, 2022, 3:33:03 AMSep 23
to

Alfred Molon

unread,
Sep 23, 2022, 12:56:34 PMSep 23
to
Am 23.09.2022 um 07:53 schrieb David Taylor:
> With the second (Monkey) image, was Topaz processing the OOC RAW data,
> or the JPEG image shown on the left?

In both cases Topaz processed the RAW files

> The Web page seems to give little information on the product - will it
> process Panasonic RAW format as well as Olympus format?

I don't know which cameras are supported (although Panasonic should be
supported). But you could give it a try.

And yes, you need a modern and powerful PC for this software 8or at
least a powerful graphics card).

RichA

unread,
Sep 24, 2022, 2:08:24 AMSep 24
to
Yes, they've really improved on this. But I would like to see shots taken in a controlled environment at 200-400 ISO then again at 6400 ISO to see how much of the reclaimed detail matches low ISO detail.

Alfred Molon

unread,
Sep 24, 2022, 5:12:00 AMSep 24
to
Am 24.09.2022 um 08:08 schrieb RichA:
> Yes, they've really improved on this. But I would like to see shots taken in a controlled environment at 200-400 ISO then again at 6400 ISO to see how much of the reclaimed detail matches low ISO detail.

Quite possible that they apply models to regenerate the detail. For
instance, they detect a monkey and apply the fur model to regenerate the
fur on the hands.
Or the AI detects a bird and applies a "feathers model" to regenerate
the detail.

Still, the regenerated detail in these images looks quite realistic.

Incubus

unread,
Sep 26, 2022, 10:40:29 AMSep 26
to
On 2022-09-22, Alfred Molon <alfred...@yahoo.com> wrote:
> Some tests with Photo AI (images shot with the Olympus E-M1 Mark II and
> III cameras):
>
> 1. The kingfisher I shot in 2019 with the Olympus 75-300 lens at 300mm /
> ISO 6400.
> Left is my RAW processing result of 2019, right the same RAW just
> re-processed with Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/kingfisher_photo_ai.png
>
> 2. A red leaf monkey I shot last month in the Danum valley (ISO 6400,
> Olympus 100-400 at 400mm).
> Left is the out-of-camera JPEG, right the output of Photo AI:
> https://www.molon.de/images/topaz_vs_dxo/red_leaf_monkey_photo_ai.png
>
> Quite impressive detail recovery, especially in the monkey picture.

It doesn't recover detail. It inserts fake detail based on what it
thinks will look okay to a human. It might be impressive but it's not
really photography in my book.

RichA

unread,
Sep 26, 2022, 8:15:35 PMSep 26
to
But at what point is basic "sharpening" in programs like PS "cross the line?"

Alfred Molon

unread,
Oct 3, 2022, 10:03:16 AMOct 3
to
It's not adding any fake detail. If you look closely at the monkey
picture for instance, you will see that Sharpen AI is not adding a
single hair which was not in the original image.

Sharpen AI seems to have some very sophiticated deblurring algorithm,
meaning that it takes a blurred image and removes the blur.

Years ago, some poster hear was using the Richardson–Lucy deconvolution
to unblur images:
https://en.wikipedia.org/wiki/Richardson%E2%80%93Lucy_deconvolution

Perhaps Topaz have done some research on the matter and found an even
better was to deblur images.

Incubus

unread,
Oct 3, 2022, 10:10:13 AMOct 3
to
On 2022-10-03, Alfred Molon <alfred...@yahoo.com> wrote:
> Am 26.09.2022 um 16:40 schrieb Incubus:
>> On 2022-09-22, Alfred Molon <alfred...@yahoo.com> wrote:
>>> Some tests with Photo AI (images shot with the Olympus E-M1 Mark II and
>>> III cameras):
>>>
>>> 1. The kingfisher I shot in 2019 with the Olympus 75-300 lens at 300mm /
>>> ISO 6400.
>>> Left is my RAW processing result of 2019, right the same RAW just
>>> re-processed with Photo AI:
>>> https://www.molon.de/images/topaz_vs_dxo/kingfisher_photo_ai.png
>>>
>>> 2. A red leaf monkey I shot last month in the Danum valley (ISO 6400,
>>> Olympus 100-400 at 400mm).
>>> Left is the out-of-camera JPEG, right the output of Photo AI:
>>> https://www.molon.de/images/topaz_vs_dxo/red_leaf_monkey_photo_ai.png
>>>
>>> Quite impressive detail recovery, especially in the monkey picture.
>>
>> It doesn't recover detail. It inserts fake detail based on what it
>> thinks will look okay to a human. It might be impressive but it's not
>> really photography in my book.
>
> It's not adding any fake detail. If you look closely at the monkey
> picture for instance, you will see that Sharpen AI is not adding a
> single hair which was not in the original image.
>
> Sharpen AI seems to have some very sophiticated deblurring algorithm,
> meaning that it takes a blurred image and removes the blur.

You cannot deblur a blurred image. You cannot recover detail that isn't
there. The application is fooling you. As I said, it might be
impressive, but the results won't always impress.

nospam

unread,
Oct 3, 2022, 11:01:02 AMOct 3
to
In article <slrntjlr9u....@localhost.localdomain>, Incubus
<u953...@gmail.com> wrote:

> You cannot deblur a blurred image.

actually, you can.

more info in the part you snipped:
<https://en.wikipedia.org/wiki/Richardson­Lucy_deconvolution>

Alfred Molon

unread,
Oct 3, 2022, 12:49:03 PMOct 3
to
Well, it seems you can - compare the images I posted - no fake detail
has been added.
And read the article about the Richardson–Lucy deconvolution I posted.

Alfred Molon

unread,
Oct 3, 2022, 1:58:01 PMOct 3
to
One more article on the matter:
https://en.wikipedia.org/wiki/Deconvolution

Combine deconvolution with knowledge how the subject should look (fur,
feather, tree bark etc.) and you should be to achieve do something.

My guess is that these algorithms will get more and more sophisticated
over time.

Whisky-dave

unread,
Oct 4, 2022, 8:34:19 AMOct 4
to
For me that's when the worrying starts is it a photo or a digital representation of what the image should look like.

An example would be a photo of Yul Brynner . I believe he had hair in the 1940s.
But if take a picture of him (or rather a picture of a photo in the mid 60s) , would the AI put hair on his head ?
.

Alfred Molon

unread,
Oct 4, 2022, 12:54:23 PMOct 4
to
Topaz Photo AI won't put hair on a bald head.

David Brooks

unread,
Oct 4, 2022, 1:42:13 PMOct 4
to
On 04/10/2022 17:54, Alfred Molon wrote:
> Am 04.10.2022 um 14:34 schrieb Whisky-dave:
>> On Monday, 3 October 2022 at 18:58:01 UTC+1, Alfred Molon wrote:
>>> Am 03.10.2022 um 18:48 schrieb Alfred Molon:
>>
>>>
>>> Combine deconvolution with knowledge how the subject should look (fur,
>>> feather, tree bark etc.) and you should be to achieve do something.
>>>
>>> My guess is that these algorithms will get more and more sophisticated
>>> over time.
>>
>> For me that's when the worrying starts is it a photo or a digital
>> representation of what the image should look like.
>>
>> An example would be a photo of Yul Brynner . I believe he had hair in
>> the 1940s.
>> But if  take a picture of him (or rather a picture of a photo in the
>> mid 60s) , would the AI put hair on his head ?
>
> Topaz Photo AI won't put hair on a bald head.

Haha!

Whisky-dave

unread,
Oct 6, 2022, 8:55:22 AMOct 6
to
On Tuesday, 4 October 2022 at 17:54:23 UTC+1, Alfred Molon wrote:
> Am 04.10.2022 um 14:34 schrieb Whisky-dave:
> > On Monday, 3 October 2022 at 18:58:01 UTC+1, Alfred Molon wrote:
> >> Am 03.10.2022 um 18:48 schrieb Alfred Molon:
> >
> >>
> >> Combine deconvolution with knowledge how the subject should look (fur,
> >> feather, tree bark etc.) and you should be to achieve do something.
> >>
> >> My guess is that these algorithms will get more and more sophisticated
> >> over time.
> >
> > For me that's when the worrying starts is it a photo or a digital representation of what the image should look like.
> >
> > An example would be a photo of Yul Brynner . I believe he had hair in the 1940s.
> > But if take a picture of him (or rather a picture of a photo in the mid 60s) , would the AI put hair on his head ?
> Topaz Photo AI won't put hair on a bald head.

What about a bald patch.

How about unbrusing a fruit.
That was my first cloning session when learning photoshop 3.0.

I'm sure AI could do that.
But when should it take over .....
We teach AI of and machine learning, it;s an interesting concept, and why the term artifical is used.
Reply all
Reply to author
Forward
0 new messages