With the revolution in AI-augmented tools, we suddenly have user-friendly software to fix many of the mistakes and augment imagery that would previously have taken a professional retoucher hours and days to fix.
Recently, I was faced with a project to restore some of my grandmother’s old photos, and I was interested in figuring out what the state of the art in face restoration and photo restoration software was.
After having looked at a lot of YouTube videos and comparison blog posts, I was surprised to find that many free and paid models were not included in people’s comparisons. My criteria were first and foremost quality, then user-friendliness and batching functionality.
In my initial research, I found the following services, software, and apps to aid me in this project.
| Name | Batch Processing | Cost | Quality |
|---|---|---|---|
| imglarger.com | no | ? | 0 |
| GPEN | possible | 0 | 5 |
| Microsoft Replicate | possible | 0 | 5 |
| Remini Mobile | no | 0 with ads | 6 |
| gfpgan 1.4 | possible | 0 | 8 |
| Remini Desktop | no | $17.99/month | 7 |
| Remini Desktop Color corrected | no | $17.99/month | 7 |
| gfpgan 1.3 | possible | 0 | 9 |
Real-world use-cases for Face and Photo restoration technology
There are many use cases for this technology. Let me highlight a few:
The salient use case for many is the fixing of AI-generated Imagery from software such as Stable Diffusion and Midjourney. With the recent advancement in AI art, we suddenly have the programmatic generation of visuals. But due to the fact that we are still in the infancy of the technology, there are multiple situations in which the output, especially for faces, is not quite where it needs to be. This is where a quick fix, photo or face restoration piece of software can help immensely.
A second use case could be, as in my example, where you have a bunch of high-res vintage photography where the faces are not quite high enough resolution for printing to super high fidelity photo books.
A third use case could be in the case of being handed low-res photography, where you need a little bit more fidelity to use the asset. This happens to me often, as I am handed photography that is super low res, or I am forced to go to Google images to find an image that inevitably does not meet the quality standards that we’re after. In this case, it has become a routine for me to feed the low-resolution photography to one of these applications or software to fix the input before the asset is used in the outputs.
A word of caution
One thing to be aware of is that while this software is amazing, it is creating pixels out of nothing. Beware of visual glitches or the software hallucinating details that were never there.
We are also skating a narrow path of truthful representation of humans. Some of the software included in this review has the potential to retouch human faces to be more beautiful and effectively not at all how they really were in reality.
I leave it up to the individual designer or artist to manage this ethical dilemma and to err on the side of caution. That being said, it is impossible for most humans to remember what the original face and photo look like, so unless you’re doing a side-by-side comparison, the majority of people would never know you had enhanced your photography through these processes.
The Original Image
imglarger.com
IMGlarger was the first software I tried. Too my surprised it seam to do nothing to my vintage photo. Maybe there was an outage when I tried. The pricing page is also down so I would not know what it would cost me to use the software on the regular.
GPEN
https://github.com/bycloudai/GPEN-colab
GPEN creates a really nice result that is unfortunately unusable due to the clear seam between the vintage photo and the AI generated image. Similar what you would see in the solution like Remini but much more pronounced. It should be possible to soften the seam
Microsoft Replicate
https://replicate.com/microsoft/bringing-old-photos-back-to-life
MS has a hard time with this photo. Sure there are no seams or hard edges and the restored image is slightly better, but the result leaves a lot to be desired.
Remini Mobile App
https://apps.apple.com/us/app/remini-ai-photo-enhancer/id1470373330
With Remini we get a relatively clean output that adds much detail to the face. We do see a very clear seam, however not as clear as GPEN for example. That being said on the right image I have known remini to be great.
Gfpgan 1.4
https://github.com/TencentARC/GFPGAN
Here we get really high Fidelity with hardly any seams, apart from the blurriness and the seam in the hair. The detail added looks natural and true to the original. I am impressed.
Gfpgan 1.3
https://github.com/TencentARC/GFPGAN
To my surprise, the older version of this technology has a slightly better quality. The area around the glasses seems more defined and believable. We still have the nasty seam in the hair which might come down to user error.
Remini Desktop
https://remini.ai/
Here we get an even better result than the mobile app. One could argue that the result blends in better than GFPgan and that it’s more believable. On a closer look, however, we see some hallucinations in the glasses area. And we also see a seam in the form of a color shift between the background sky and the area that has been cut out to be augmented.
Remini Desktop Color corrected
https://remini.ai/
As a bonus, I am including a photo of the automatically enhanced background base and color-correct model from Remini. This is a solid brightening up of the slightly under-exposed original photo and something that might be valuable to you if your colors are messed up.
Conclusion
As a paid option, Remini is truly great at giving you a quick and usable result. There is a tendency to hallucinate details that were never there, and thus manual review of each photo that you send through the process is necessary. Additionally, batch processing is not yet available, and thus it is a tedious process to process a full folder of photography.
I was able to make it a little bit more tolerable by creating a Keyboard Maestro macro to cycle through, upload and download the photography from Remini’s web version, but even then, it’s not ideal.
The free solution GFPGAN comes out on top on this image however — and the fact that it is free makes it that much more useable.
This software is bound to just keep on getting better. And with that, we’re going to see a lot of ethical issues arise when the truth and the imagery are too far apart. I’m reminded of the recent controversy over the new TikTok filters that are deemed too good and too untruthful. I urge designers and artists to use good judgment when using these tools.
