Topaz Photo AI v0.9

If you are in a mental dead end as to what exactly is going on here, it is worth reading now and then (ok actually always) what TL actually wants to achieve.

I understand that you build it (autopilot) first because it has to be able to handle the other apps.

1 Like

@01matthunter
I would definitely check your firewall disable with Tinywall it was my breakthrough, AV was still on.
Make sure that you turn it back onn though.

@yazi.saradest

False Detection:
Subject was selected correctly.
The butterfly is out of focus, that is correct.

But the noise is high, this should be removed.

Although Sharpen AI from PHAI is capable of denoising I can’t judge that here, the Sharpen version must be very old.

Uploaded along with the other photos from the same series with the butterfly, 35 images.

Zip Name: Butterfly_TPX

Checksum:
SHA256: 4534238C16249193149DDDD6089B1F004EE744018B97E7F8F91F3A14E0D4D378

1 Like

I just tried to do “Save Image As” in my browser, only to find that the JPG file type had been hijacked in the files types pulldown list by “Topaz Photo AI 0.8.0” !! Please give me back my JPG option! (I checked – it did save as a JPG – but I don’t think it fair that Topaz should lay claim to all JPS under Windows! :blush: )

I noticed (at Photo AI Alpha phase), that the Natural upscaling model was a lot more “natural” - it didn’t sharpen the image (Sharpening part of the Photo is OFF, of course). Is it possible to re-add the old Natural model under a different name? It did great for the images with an intended touch of chromatic aberration. The current “Natural” model sharpens the edges and practically removes the effect.

An example of OLD Natural model and Sharpening OFF (Alpha didn’t have Face Refinement ability back then, right?):

I also noticed, that the Face Recovery “elements” (eyes, mouth, etc.), while the FR number is high, tend to be of lower quality, than the original ones, when the face elements are big (like a purposefully created anime-like character in CG). When the FR number is low - it’s perfectly fine.

Both following examples are 4x Resize’d, with Natural model, Sharpen is OFF, preview at 100%.

FR at 32:
2022-09-04_17h14_59

FR at 99:
2022-09-04_17h15_56

Sharpen (Motion Blur, strength at 41), subject only, Face Recovery at 23 (FR at 99 didn’t give any noticeable change in sharpening), Resize 4x with Natural model:
2022-09-04_17h22_15

It seems it’s better to to turn the Sharpen OFF in this example (though, Autopilot turned it ON, with parameters I described up here).

Is there a way please to crop an image before rezising (topaz Photo ai standalone)?

Did A.P. turn on Denoise in this test of yours?

In my pics, Denoise only (Sharpen at zero) does a fair bit of sharpening vs just denoising. If I feel the need to try to sharpen more with the sharpening slider it starts mutating my image & creating halos w/out necessarily naturally crisping specific elements I want sharper.

I’d prefer that denoise just denoise. And sharpen just sharpen in a multi-functional integrated product.

Intuitively, it seems (to me) that the functionality overlaps make it more arithmetically complicated for an AutoPilot to calculate proper balances…

It’s like if a chef had to make a recipe that called for zero salt, but all his/her ingredients contained a noticeable amt of salt.

Win 11. PAI .9 Ps 2022 (23.5 rel.) Plugin tests.

I’ve been recording Ps ‘how to’ videos for YouTube for a local photo club I used to attend/give talks at (pre-pandemic, when they met). I’m currently working on a video re: Ps “Selections” methods & techniques (automatic ones as the 1st in a series).

Because I’m doing this, I wanted to compare my Ps auto-select Subject results (I’m also trying out Ps’ Cloud-based vs on-device subj. selections) with those of PAI .9 Ps Plugin; in case I should feature Topaz PAI in my video wrt the Ps use of it as a plugin…

I purposely chose images (for a portion of the video) that were more challenging in terms of subject close to background tonality to explain how to deal with that. Not every image someone will have to process will be a tidy clean portrait on an 18% grey or white uncluttered backdrop.

Here’s what I found with 2 images I’m considering using in my how to video:

Man & Dog (Ps Subj Select - cloud-based vs Ps Plugin PAI .9) - labelled in filenames: (BTW, PAI detected the man’s face)

Woman & Cat (Ps Subj Select - cloud-based vs Ps Plugin PAI .9) - labelled in filenames: (BTW, PAI didn’t detect woman’s face - I’ve already Dropboxed this image previously)

I will Dropbox the photo of the Black man with his cute blk & wht dog.

I have just installed Photo AI 0.9 on my MBP M1Pro and am impressed with the concept and it seems to work well. I did notice that if I imported a Fuji RAW file (RAF) it is cropped on all sides by Photo AI. I imported identical photos - one was the original RAW file, and one was a TIFF created in Lightroom. The TIFF image was processed in its entirety but the RAW file shows up as cropped as soon as it comes into Photo AI. Photo Below.


However it seems to work fine if I go directly into Photo AI from Lightroom.
Thanks
Tom

No faces eh? Not even a dog? :wink:

What an online service did with the same image:

Enhanced_photo-3

And I wasn’t able to turn on Face Recovery manually to force it.

I was going to test GAI in comparison (M1 MBP, Ventura) but it’s now crashing upon loading the image.

GAI on Intel (2013 MBP, Monterey) does not crash. Just makes the fans throw a fit…

Aha! That’s a relief.

And again!

A severe crop helped PAI along…

1 Like

Yes, recently standalone Gigapixel AI 6.0.0 on my M1 Mac Mini 16gb running MacOS 12.5.1 sometimes crashes when it loads an image.

OK, so we have GAI crashes on Monterey and Ventura on M1, and it’s happening with 2 different builds of GAI (6.0.0 and 6.2.0). But not on Intel (6.2.0).

And yes, I always use the apps as standalone.

Hmm!

Example of subject detection – almost!

The extreme crop made a huge difference! But a lot of images we can’t crop like that to get a result like that …

Yes, it did, on previous beta of Photo AI (v0.7 and, I think, v0.8 too), but not on current one (v0.9).
I noticed various functions overlapping themselves - sharpening mostly (with Resize it might have something to do with the model, therefore my question about including older “Natural” as well)

I still think PAI needs to add Auto Color/Levels to its Autopilot, otherwise you need to take your noise-reduced, sharpened, resized image into some other app to finish the job.

This can be added as a simple On/Off slider like the rest, with no other settings to keep it simple. Or it can happen under the mysterious Enhance setting…

In this sharpened and resized example I think the white level needs to be popped just a bit, but PAI doesn’t offer the option:

There’s an upcoming app I have been working with that automatically fixes tone and color in one shot, and I think this basic corrective tech is perfectly suited for PAI at the very least (since all of the fun and useful Topaz Studio stuff has been removed, and most other developers still retain that sort of thing in their all-in-one’s).

Examples of the low-level correction I would like to see as part of the PAI Autopilot:

Speaking of Topaz Studio (if I may wax nostalgic), today I confirmed it runs fine even in macOS Ventura (beta 6). My original iPhone freeway grabshot, and some TS effects:

4 Likes

In a more modest way than the scope of Studio (which is great), Classic Topaz Adjust also provides features to add ‘pop’ and finish beyond basic denoising & sharpening (of any subject - face possessing or not…).

I almost mentioned Adjust! That’s how Topaz got on the map some 15 years ago.

Yeah, how about an Adjust slider in PAI! Or rename Enhance and make it do the same thing…

Yes. Then everybody wouldn’t be asking, “What the [heck] does Enhance do?”

But there must be very fast responsiveness to seeing the effects of such a slider’s (or set of sliders’) movement…

One of Adjust’s presets was called Photo Pop. If it was used as a strength continuum it would play with the design concept in PAI of converging on a single model like Std too.

“Subject detected” does a lousy job