Today, with the rapid development of large language models and various AI applications, I finally made up my mind to digitize all the paper film photos in the physical albums of the previous generation before the old photos are completely yellowed, faded, dampened, or damaged. , and perform a thorough AI restoration so that these precious memories can be preserved for a longer period of time.
The way to digitize is to use the "Photoscan" app provided by Google Photos, which automatically adjusts angles, crops edges, eliminates glare, shadows, and saves both the original and the automatically edited pictures in iCloud Albums. (It is more recommended to use a scanner for maximum resolution, if possible.)
The next problem encountered is how to choose a suitable AI repair tool. Based on past experience and recent popularity, I have selected 5 candidates. In order to achieve the best repair effect, I tried almost every one of the hundreds of old photos to be repaired with all the tools, which can be regarded as an in-depth use of these tools, so I will share the experience with you below .
tool list
| tool | platform | price | Model |
|---|---|---|---|
| CodeFormer | Web page | free (open source) | Codebook Lookup TransFormer (CodeFormer), 2022 年 |
| RestorePhotos.io | Web page | free (open source) | GFP-GAN, 2021 |
| jpgHD.com | Web/iOS/Android/Mac/Windows | Charges, 5 free trials available | Unpublished AI super-resolution model, 2021 |
| you and me | iOS / Android | Charges, watching 20 seconds of ads can process 1 | unpublished |
| old photo restoration | iOS / Android | TOLL | unpublished |
Here are two additional open source tools:
- CodeFormer ( official website | GitHub project address ): The work of Zhou Shangchen and others from Nanyang Technological University, based on the Codebook Lookup TransFormer model. You can use the demo built by the author through the online model running service Replicate to experience directly (limited by the free quota of Replicate, and you need to pay if you exceed it). Attach the parameters I use for reference:
fidelity = 0.8,background_enhance = true,face_upsample = true,upscale = 2 - RestorePhotos.io ( official website | GitHub project address ): An old photo restoration tool based on the GFP-GAN model of Tencent ARC Lab. The official website can be used 5 times a day for free, but it can also be deployed with one click through Vercel, etc., and fill in your own Replicate API key for unlimited use.
Note: Due to understandable privacy reasons, the photos I actually used for restoration will not be shown below. I collected some pictures from public sources for testing and for your reference.
The first set of test pictures: single and double (big photo)
Test 1-1 Teresa Teng
This is Teresa Teng in a screenshot of a TV show. Almost all tools have restored this picture well, including hair and earrings ("CodeFormer" has some flaws in the detail processing of teeth). Which picture do you think is more like the real Teresa Teng?
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
Test 1-2 Jay Chou
The original picture of this Jay Chou childhood photo is blurry and has a lot of noise. From the perspective of restoration performance, "CodeFormer", "You and Me Back then", and "Old Photo Restoration" are more excellent in background noise reduction and clothing texture restoration, but only "jpgHD.com" accurately reproduced Jay Chou's single eyelid and interior. The details of the corners of the eyes, in my opinion, are the most similar to the original appearance of the picture. Other tools are appropriate "imagination".
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
Test 1-3 Photo of Leslie Cheung and Cecilia Cheung
A blurred group photo, the repair effect of the five tools can be described as consistent.
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
Test 1-4 Group photo of Jacky Cheung and Joey Wang
Except for "jpgHD.com" which handles Wang Zuxian's skin rougher, each tool is almost equal in the processing of this kind of single-person and double-person headshots. Note that all tools restore the earrings.
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
Test 1-5 Einstein
Here is an attempt to restore the three great gods Albert Einstein, Hendrik Lorenz, and Marie Curie at the Fifth Solvay Conference in 1927. The original image is black and white, and has been artificially colored, but the details have not been enlarged and repaired. It can be seen that "CodeFormer" has better repair effects on faces and suits, and successfully recognized Arthur Compton with only half a face behind.
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
Test 1-6 Lorenz
Note that Old Photo Restoration incorrectly generated a blurry outline of the glasses.
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
Test 1-7 Marie Curie
In terms of skin texture, the clarity of "CodeFormer" is ahead of other tools.
Left: original image, right: CodeFormer
Left: RestorePhotos.io, Right: jpgHD.com
Left: You and me back then, Right: Restoration of old photos
The second set of test pictures: multiple people (big group photo)
Test 2-1: Group photo with slight blur
The author of the picture is Wu Kao, and the original title is "Under the Lens: Chinese Children in the 1980s". ( source link )
original image
CodeFormer
RestorePhotos.io
jpgHD.com
you and me
old photo restoration
Test 2-2: Group photo of four people with moderate blur
Photo source: Campus photos taken by students of the International Trade and Economics Double Degree Program of the School of Economics and Management, Nanjing University of Science and Technology (summer 1990, from left: Han Mengling, Zhou Xiangwen, Shishi, Lu Hong). ( source link )
The processing of this picture has produced a more obvious difference. "CodeFormer" has a more natural face, successfully restored the glasses of the third girl on the left (but produced burrs), and repaired hair, clothing and buildings more accurately; " The face restoration effect of "You and Me Back then" is very natural, and "Old Photo Restoration" also performs very well in building restoration. However, "RestorePhotos.io" has a large degree of distortion in facial restoration, and the glasses of the third girl on the left have not been successfully repaired.
original image
CodeFormer
RestorePhotos.io
jpgHD.com
you and me
old photo restoration
Test 2-3: Highly blurred group photos
Photo source: Group photo in front of the small hall of the West Building of Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences in the 1980s (from left, front row: Wang Huimin, Zhang Lianying, Yang Hengcai, Yu Abin, Gu Deying, Li Ailian, back row: Lin Yueming, Fang Zujie, Zhang Weizai , Jin Zhiliang, Liu Zhiwu, Hu Yanzhi). ( source link )
The face in this picture is blurred to a large extent, and each tool has given full play to imagination for brain supplementation.
From the point of view of the natural harmony of the facial features, I will sort the restoration effects in this way: CodeFormer > Old Photo Restoration > jpgHD.com > You and Me Back then > RestorePhotos.io.
From the restoration effect of the background and clothing, I will sort the restoration effect as follows: CodeFormer = Old Photo Restoration > You and Me Back then > RestorePhotos.io = jpgHD.com.
original image
CodeFormer
RestorePhotos.io
jpgHD.com
you and me
old photo restoration
test conclusion
| tool | Quality improvement | Clarity | Accuracy | repair speed |
|---|---|---|---|---|
| CodeFormer | Support optional image quality improvement (custom magnification) |
very high | High Compared with the original image, the accuracy is higher, but there may be a small amount of unnatural details |
Generally ranging from 6s to 60s |
| RestorePhotos.io | not support | The maximum resolution of the generated image is 1200px |
Low Compared with the original picture, the degree of "brain supplement imagination" is greater, and the performance on small avatars is relatively poor, and it is easy to be distorted and deformed |
Very fast Usually can be completed within 10s |
| jpgHD.com | Support optional image quality improvement (2x, 4x) |
medium | Very high Can restore the real details of the original photo |
Very slow , even small images can take 3-5 minutes |
| you and me | Support default 2x |
higher | Medium May appear distorted for small portraits |
Soon, usually around 10s |
| old photo restoration | Support default 2x |
higher | higher | Soon, usually around 10s |
On the whole, I recommend the free and open source "CodeFormer". If you have the same need to repair hundreds of photos at the same time as me, you can build "CodeFormer" locally, which will save you a lot of worry. The performance of "You and I Back then" and "Old Photo Restoration" are fairly stable, but it is not as good as "CodeFormer" in the performance of small avatars, and it is a paid software.
Tools other than "RestorePhotos.io" support coloring of old photos, and this feature is not tested in this article. "RestorePhotos.io" has certain advantages for headshots and difficult image processing with more frosted particles or interference fringes, but it is not ideal for other situations.
epilogue
Due to limited energy, this article only tests and compares some common situations, hoping to help you make a preliminary selection judgment. In fact, there are still many difficult and miscellaneous problems in the process of restoring old photos, such as particles caused by frosting, interference fringes caused by re-shooting screens, and some problems such as smearing, creases, tears, and mildew. When facing these problems, you can try different tools for comprehensive comparison, and even more patient manual intervention is required to superimpose the effects of different repair tools.
After the initial restoration of old photos, you can also try to use these tools for further lossless enlargement and intelligent coloring. It is also recommended that friends with strong hands-on ability do further post-processing on those very important and precious photos, such as artificial color correction and texture. After all, although AI can already help us restore the original appearance of the photos to a large extent, it cannot reproduce them 100% accurately history. For friends who have extreme requirements for photo restoration, I suggest you contact a team called "TR Art", and I will attach a comparison picture of Zeng Guofan and Zuo Zongtang manually restored by them for your reference.
Restoration of Zeng Guofan and Zuo Zongtang's portraits by "TR Art"