About HitPaw Photo Enhancer AI

[lhkjacky (Jacky)]

There were some discussion about this new AI software for Photo Enhancement. @itunes663

@itunes663

After some testing, I would like to summarize what I have found.

---

Cons:

1. HPE only upscale in 4x.
   There is no way to input scale factor or dimension.

2. There are 3 AI models available, but there is no way to adjust the level of noise reduction or sharpness.

3. The General Model is good for landscape. The Denoise Model is good for CG Anime. The Face Model is good for Portrait.
   It will be challenging if a photo contain both person & landscape. Using Face Model, enhance the face nicely but the background detail will be all removed. Using the General Model, the background upscale nicely but the face will be distorted.

4. No way to select export image format or compression.
   (For JPG, It always output at 100% quality).

5. No Raw support (only JPG, PNG, Webp)

6. No auto Preview, you have click the Preview button every time and it will process the whole picture.

7. At the moment, only Windows version. (There will be Mac version but ETA unknown.)

---

Pros:

1. Interface is very simple to use.

2. The Face Model generally give excellent result on tiny face.
   I suspect the face is upscale with GFPGAN, while the background is upscale with Real-ESRGAN ?
   When the person is looking strange to the camera, it usually restore the face detail nicely. But in extreme case, such as a large group of kids running around with motion blur in a small photo, where the faces are barely recognizable, the restoration may create monster face.

---

I will update this post later if I find out more.

[itunes663]

Sounds about right. Although I find myself using HitPaw Photo Enhancer AI More often than I expected, actually more often than Topaz now. I find it to work really well with tiny faces and with combination of Photoshop compositing it can take old photos or blurry photos or small faces from smartphone and reconstruct them in ways I cannot do with other software. It might be something one can do with these AI models you mentioned, but usually it requires some skills to set it up, and HitPaw AI uses simple to use interface and also works in batches.

So I tend to use it for faces and either Model A for everything and Model C for faces to create overall better images. It also does magic with enlarging small reference photos from the internet and screenshots of all movies. I sometimes use it as reference for paintings.

Anyway, why does Topaz not do the same if the AI model is available for training, I don’t know. But Topaz should include it in their products.

Here is an old painting that I could never find sharp photograph on, anywhere online. So I was really pleased how it was able to “un blur” it. something I could not even do this well with Topaz Sharpen AI.

[lhkjacky (Jacky)]

Currently, GAI have a new beta (v5.8.0.3b) which improve tiny face upscale.

@alexander.zhang

This update introduces an overhauled small face refinement pipeline, which should both detect human faces more often, and deliver more realistic results. Tiny background faces especially should see a noticeable improvement in quality. Refinement has also been moved to the GPU where possible, making images process much faster. Note that large faces are not affected by the face refinement switch, as the regular models still deliver better results.

The Face Enhancement in V5.8 beta is much better than V5.7.3 but it is still far behind GFPGAN. (HitPaw face restoration is identical to GFPGAN)

@itunes663

Anyway, why does Topaz not do the same if the AI model is available for training, I don’t know. But Topaz should include it in their products.

I have asked the same question in the beta thread, developer said, they are working on it right now.

[itunes663]

Oh, well that is good news. Looking forward to testing it, although I find Topaz releases often very buggy version even when they claim its not a beta, so I treat beta as alpha and I’ll wait for a while to see if its ok to use. Its a large download after all, but its good news.

I wonder can GFPGAN be also trained on animal faces as well. I know the database of such images is smaller and animal species are numerous, but I would assume its something that will see in the future. Similar how camera manufactures started to detect human eyes for auto focus and later for animal and pet photography they started adding animal eye auto focus, because every camera manufacturer was already good at humans. I suspect that is the next race in the AI photo enlargement story. After humans are something everyone can do, next up, reconstruction of tiny animal faces.

[itunes663]

Here are two more tests. One is the demo image of a blurred face that ships with the program for demonstration of that it can recover and the other is dirty, low light, shaky test of cheap smartphone selfie cam to see how much it can recover. It is pretty impressive how much it can recover, I must say. And with those that have retouching skills in Photoshop its possible to get much better results with skin pores added or at least film grain to give back some texture and other things, even re-lighting the shot with other tools.

Retouch4me, AI based retouching plugins - Retouch4me Is Truly Phenomenal. A set of plugins I’m truly impressed with, the results I get are incredible. It’s crazy how the plug-ins reduced my retouch time by 70 to 80%. I use to spend hours and hours on each photo and my entire retouching workflow has completely changed overnight.

Retouch4me — Advantages of Artificial Intelligence (AI) in Retouching

I must say I think Topaz is starting to lag behind. they had a head start, but competition is catching up or surpassing them in some areas.

[plugsnpixels]

lhkjacky, thanks for splitting off this discussion! First I’ll repeat my report about Photo Enhancer from the other thread and continue from there:

"I figured I’d give this a try as I’m very happy with Gigapixel and use it to salvage either my older digital images and film scans and to upscale smaller AI app-generated art.

"First of all, HitPaw’s app doesn’t work under VM (Windows on a Mac) which you find out after you install and try to run it, and as you said there is no Mac version. So I had to fire up a PC and start again.

"On the PC, I still don’t know how good the app is because it refuses to generate a preview (“0%” flickers when I hit the Preview button but nothing happens). I also can’t see the two before-and-after panels you show (I see a single panel) or export the result to view separately when using the trial version. So that’s the end of that.

“The average consumer is not going to mess with ESRGAN! “esrgan requires python >= 3.8. The requirements.txt file can be used to install the necessary packages.” Check the text file: ‘This page does not exist yet.’ OK…”

I hesitate to use their included examples as I don’t know whether they intentionally blurred the “before” and use the actual image for the “after”. I need to use my own. At any rate, the difficulty getting it working properly speaks to some quality control issue and makes it feel “cheap”. I can try one other PC and see if I have better luck.

The previous best facial reconstruction technology I tried came from MyHeritage and whatever technology they licensed. That seems to be on par with the Photo Enhancer examples. I brought this up to the Topaz developers a few years ago.

A couple of examples from MyHeritage, the first involving my great-great grandfather, sourced from a scan of the 5x7 original (he was part of a distant group shot):

I was very impressed with that!

From the same original photo, his son, my great-grandfather:

Nice try, but in this case, MyHeritage seemed to reconstruct the facial data from other sources, and not well. Which reminds me:

GAI early on seemed to use a database of face parts that I called “Mr. Potato Head”. A few years ago I did a blog post where I showed examples of GAI facial enhancement using actual data and, at the end, an example of the “Mr. Potato Head” restoration of a distant face where there wasn’t enough detail to enhance the real features.

It’s a tough call in those cases and I’m not sure what the proper answer is.

But if GAI can get to the point where it meets or exceeds the results from MyHeritage and (supposedly) Photo Enhancer, we will be very happy.

[itunes663]

Very cool. I hope you don’t mind, I have HitPaw AI a try with the images you presented here, but I think because they are already cropped it could only reconstruct the facial data. I would like to try with full image if you want me to.

Does MyHeritage uses online processing, because for me I really aim at mostly offline processing, for privacy, security and efficiency.

[itunes663]

"“On the PC, I still don’t know how good the app is because it refuses to generate a preview (“0%” flickers when I hit the Preview button but nothing happens). I also can’t see the two before-and-after panels you show (I see a single panel) or export the result to view separately when using the trial version. So that’s the end of that.”

Hmm I can’t really speak as to what the problem might be, but sometimes it takes a few moments for it to process the image in initial stage if its a very large or complex image. The flicker thing is just a progress animation they added while you wait. Once its done you get the two side by side preview windows and you can export your fail. When working in batch mode, you only choose a folder where you want to export and what model you want to use.

I think they demonstrate it in this video, but it was release before the new Face Enhancement Model was added.

[2022] Best AI Image Enlarger? Top 4 Image Enlarger Compared, Better or Worse?

[2022] Best AI Image Enlarger? Top 4 Image Enlarger Compared, Better or Worse?

[plugsnpixels]

Thanks for your examples! I PM’d you the original.

I’ll give the app another try and see if it makes any difference. I agree about offline processing, though for some scenarios an external render farm might be more useful.

[itunes663]

Got your message. I’ll post my results a bit later. Thanks.

[plugsnpixels]

Tried to install the app remotely on a work PC, needs elevated privileges (I am an admin…).

It’s becoming a PITA, I will wait for the Mac version.

[itunes663]

I used Gigapixel on the rocks because it did a slightly better job than HitPaw, but I used HitPaw for faces where it does the best job. Than in Photoshop I simply do simply masking to glue in the faces as it were and after that I applied a bit of blur and some film grain and contrast adjustment to make it all match a bit better. And as a final touch I did here and there a light touch up using heal and clone tools to remove the most distracting scanning debris. The image really needs to be properly retouched manually to be restored, but you can see how it could be done. I think., considering the starting position its pretty impressive what these tools can do. This is by the way just some crop of people in the picture, 4x enlarged from the original , here shown at 100% magnification.

[jgr00ms2 (m4/3 photographer)]

@itunes663

Retouch4me Is Truly Phenomenal.

At that price, one would hope so…$1000 for 8 plugins (approx $125 each).

[itunes663]

If you are working professional and can see that as investment, it definitely saves you lot of time. For example dodge and burn manually for a face would take between 2-6 h for experienced retoucher. With this tool its 5 min. And results are virtually identical to a client.

A retoucher might not need all the plug ins. Also if you are working with studio images, cleaning up the floor automatically in batch saves a lot of time and with a few jobs you could make up for the entire collection much less one of the plug ins.

That being said, I agree it is expensive and its not for everyone or even every retoucher. For example there are some of the plug ins for removing red blood vessels in the eyes. I personally rarely have to do that, and its not a problem to do manually, so I would not buy that one. Someone who does a lot of beauty work and close ups or portrait headshots might need it more etc.

[plugsnpixels]

Thanks for your work on this! Looks very good considering the original, though my great-grandfather grew some glasses

If you see the Mac version pop up, please alert me!

[itunes663]

Hehehe. We can always add or remove glasses., hehehe.

Sure, I’ll post probably here if I see Mac version. But overall I think its a company worth following, they are releasing about half a dozen products and I really like their approach and how they are trying to approach it. You can sense the company is more than just a cash grab operation. I have a feeling in the upcoming years they might be releasing some really good stuff.

Cheers!

[lhkjacky (Jacky)]

If you are planning to use the face enhancement only, you can just use Tencent ARC’s GFPGAN. You can download the code from Github and it is free.

From what I have tested, HitPaw Face enhancement is identical to GFPGAN.

Gigapixel Face Refinement vs Remini

You have to install Python in your computer and clone all the files to your computer. Then install all the required stuffs listed in the requirements.txt. https://github.com/TencentARC/GFPGAN/blob/master/requirements.txt Then download their latest pre-trained AI model from Google Drive. https://drive.google.com/drive/folders/17rLiFzcUMoQuhLnptDsKolegHWwJOnHu?usp=sharing After all the stuff installed correctly, you can use Command Prompt to run the script. Such as, python inference_gfpgan.…

@lhkjacky

Tencent ARC’s GFPGAN:

[topazsharpenai (Splitty)]

I am not convinced that those tools are actually offline.

The HitPaw photo enhancer actually says “upload” when you add a photo, so it might get sent somewhere.

Also, never forget, if you are using something for free, the service usually sucks up all your data including your images. I surely won’t be using anything Tencent touches, they are basically a part of the Chinese government.

[itunes663]

Where does it say “upload”? I am pretty sure its not online, since I’ve done it the same way when internet is not activated. Actually their whole sales pitch is that they are not as much better than other tools , but they are offline and indeed as far as I can tell its offline. Unlike most online tools. That is what I don’t like about the other tools. Privacy concerns, slow and not very secure.

[itunes663]

Tencent ARC’s GFPGA does look almost identical to HitPaw AI and what they are using. Makes me feel uneasy when I realize that since Tencent is essentially CCP (Chinese Communist Party) and they use mass surveillance on everyone, and that is how they are able to reconstruct faces. Once you realize that, its well, a lot like sausages… its best not to know how its made.