V2.0.1 faces look like wax museum

I did a test run of v2.0.1 with 60 year old color photo of 3 people, on autopilot their faces looked very waxy, like museum statues. the 3 were playing guitars, unfortunately the software doesn’t recognize that hands are part of the body, it would be great if Photo AI2 could recognize them as well, as you can see they were essentially not recovered and look unnaturally blurry

1 Like

Can you upload the original image file as well? I want to see what it looks like to recommend different workflows or filters that are different from what Autopilot suggested.

Hi Tim

Here are the before and after photos. I used this as a test subject after I updated to v2 because it represented a vintage colorized photo, which is often difficult to work with because early colorizer artists usually painted over facial details. As you can see my results were very pasty, I compared it to a figure in a wax museum. However my specific issues isn’t the faces, which I expected to be a major challenge, but the hands, which in the original looked relatively normal and my expectation was that they would be processed using a similar algorithm as the faces. But the actual result was that the hands became blurry and a bit grotesque and if this had been an actual client’s photo I think they would have been disappointed. What are your thoughts on the result? Am I expecting too much?

Thanks
Jerry Marcin

IMG_3953.JPG

Unfortunately the hands are not included in the Face Recovery filter and are not processed with a similar algorithm. The AI models rely on existing image data so if the original image lacks detail, the program will struggle as well. That’s what we see happening here.

Try turning on the Upscale filter which will enhance the resolution of the entire image rather than turning on Sharpen or Face Recovery. Let me know what you think of those results.

Will do, I’ll play around with the filters and let you know. Photo AI has become a far more important tool for me than I ever imagined. It’s hard to convey to you the reaction I get when I restore a vintage photo and barely recognizable blurry faces suddenly become real people with details in their expressions that were impossible to previously see. Most people don’t care if non-animate objects in the photo have detail restored but the issues with hands and other exposed skin can be a turnoff. Other photos I’m working on are resulting in almost magical face restoration but exposed arms and legs (and hands) actually get worse, taking on a shriveled up appearance. Bottom line, Tim, is the ideal is if Photo AI could recognize and apply the same processes as it does for faces. I’ll try to send you some extreme examples of your app actually making arms and legs worse rather than better.

PS - you may recall I opened a ticket soon after I bought Video AI. I’ve followed all of Topaz’s suggestions and I have not once had a successful result. I’ve now completely reset my high-end workstation to ground zero and will try with a fresh install but to date I have had zero luck trying to use Video AI and I’m pretty discouraged.

Jerry

Sent using the mobile mail app

1 Like

I agree with you. Hands, fingers are one of the major weakness of Topaz Photo AI.

Here is an example.
Source:
a2c47bb07b3e683ea26007d7734e29c1459effbe_cr2

Upscale 4x:
Left: TPAI Low Resolution (Default Auto Pilot)
Middle: TPAI Standard Model
Right: Others


Source:
a2c47bb07b3e683ea26007d7734e29c1459effbe_cr

Upscale 4x:
Left: TPAI Low Resolution (Default Auto Pilot)
Middle: TPAI Standard Model
Right: Others


Source:
a2c47bb07b3e683ea26007d7734e29c1459effbe_cr4

Upscale 4x:
Left: TPAI Face Enhance
Right: Others

1 Like

I have found that the addition of a slight amount of noise to waxy faces (in post processing) can help, it would be nice to have that as a preference option in Topaz AI V2.0.1
I generally keep the strength of face recover as low as possible generally under 50 if the image works:

Great advice thanks. I see a very real difference, the lips look real, the area under his eyes is improved, basically the bottom line is that it is a 70 year old photo, it looks most real if there are slight imperfections. In your example he looks human, not wax, but the photo still looks vintage.

1 Like

I’ve studied your examples for an hour, you have created excellent examples that Topaz needs to study as well. Taking the first example, the closer view, I don’t know what the “other” process is but those fingers in a 70 year old photo look totally natural. The Topaz low res mode is really bad (but I’ve got some that are worse) and their standard mode still looks blurry. It’s interesting how Photo AI sharpens the fret of the guitar but is totally confused by the human hand, it’s not even waxy like the face. The “others” image manages a realistic balance between sharpness in the guitar and a very pleasing and acceptable hand. Frankly (to Topaz) I couldn’t present either of the first two as representative of my work. But the third example does prove that it can be done.

1 Like

The “Others” images are upscaled using the latest AI upscaling algorithm called DiffBIR, developed by the Chinese Academy of Sciences and Shanghai AI Laboratory. They recently released their research paper in arXiv.

DiffBIR utilizes stable diffusion and controlnet to upscale images and regenerate realistic details, including hair, skin texture, clothing texture, text, trees leaves, grasses and more.

Based on my testing, their face recovery surpasses that of GPEN, GFPGAN, CodeFormer, and VQFR, providing more detailed and natural-looking results.

Here is an example from their website.


Few months ago, in the June Photo Roadmap,
Topaz’s CEO Eric has asked us opinion about using Generative AI to enhance image.

Within that topic, I have replied that

Now, It seem that DiffBIR has exactly done what I suggested. :yum:

However, Topaz seem to have decided to add option 2 “remove object” to TPAI V2.1 instead. :disappointed_relieved:
In my personal opinion, Topaz should focus on improve upscaling and image quality at first piority.
Remove Object can already be done extremely well by using open-source free software such as, Lama Cleaner.

5 Likes

This is really incredible! :grinning: :+1:

3 Likes

Upscaling Quality should be the absolutely first Priority!
DiffBIR is incredibly and Adobe also working on GigaGAN :slight_smile:
Therefore, Topaz should catch up with the competitors here :wink: :+1:

3 Likes

Our developers are also working on the Upscale models to improve the quality of realistic details!

4 Likes

I can see that there’s “glasses” that are being added to the nose bridge. I’ve had a few reports of similar artifacts.

1 Like

It makes sense why the hands and other body parts matters for pictures like this. We’ve just focused on prioritizing faces in the time being due to limited capacity.

Also, if you try Video AI and it doesn’t work still, I’d recommend submitting another support ticket because it’s a different team member who manages that app so I wouldn’t know much about it.

Time flies, and it’s hard to believe that 7 months have already passed! :alarm_clock:

Do you have any new information about “Upscale models to improve the quality of realistic details!”? :thinking:

Within the last 6 months, many online free open-source upscale AI algorithms have made significant improvements. However, it appears that TPAI’s AI upscale model hasn’t made any major advancements in enhancing realistic details such as hands, fingers, and so on.

Source Image:
source image

Upscaled by free AI algorithm.

Source

3 Likes

Hi. We recently introduced Recovery model (beta) - please find here and roadmap. It should be noticeable faster than the one you mentioned - the approach we took considers the computation cost on average gpus.


2 Likes

Thank you for your reply.

Here is a full-sized comparison between the Recovery model vs other AI. :grinning:

Which other AI Software do you use? The result is incredible :wink::+1:

I have sent you a private message. :grinning:

1 Like