A Deep Learning based project for colorizing and restoring old images (and video!)
A Deep Learning based project for colorizing and restoring old images (and video!)
This Reposisitory is Archived This project was a wild ride since I started it back in 2018. 6 years ago as of this writing (October 19, 2024)!. It's time for me to move on and put this repo in the archives as I simply don't have the time to attend to it anymore, and frankly it's ancient as far as deep-learning projects go at this point! ~Jason
Quick Start: The easiest way to colorize images using open source DeOldify (for free!) is here: DeOldify Image Colorization on DeepAI
Desktop: Want to run open source DeOldify for photos and videos on the desktop?
In Browser (new!) Check out this Onnx-based in browser implementation: https://github.com/akbartus/DeOldify-on-Browser
The most advanced version of DeOldify image colorization is available here, exclusively. Try a few images for free! MyHeritage In Color
Having trouble with the default image colorizer, aka "artistic"? Try the "stable" one below. It generally won't produce colors that are as interesting as "artistic", but the glitches are noticeably reduced.
Instructions on how to use the Colabs above have been kindly provided in video tutorial form by Old Ireland in Colour's John Breslin. It's great! Click video image below to watch. Get more updates on Twitter .
Simply put, the mission of this project is to colorize and restore old images and film footage. We'll get into the details in a bit, but first let's see some pretty pictures and videos!
Note: Click images to watch
"Migrant Mother" by Dorothea Lange (1936)
Woman relaxing in her livingroom in Sweden (1920)
"Toffs and Toughs" by Jimmy Sime (1937)
Thanksgiving Maskers (1911)
Glen Echo Madame Careta Gypsy Camp in Maryland (1925)
"Mr. and Mrs. Lemuel Smith and their younger children in their farm house, Carroll County, Georgia." (1941)
"Building the Golden Gate Bridge" (est 1937)
Note: What you might be wondering is while this render looks cool, are the colors accurate? The original photo certainly makes it look like the towers of the bridge could be white. We looked into this and it turns out the answer is no - the towers were already covered in red primer by this time. So that's something to keep in mind- historical accuracy remains a huge challenge!
"Terrasse de café, Paris" (1925)
Norwegian Bride (est late 1890s)
Zitkála-Šá (Lakota: Red Bird), also known as Gertrude Simmons Bonnin (1898)
Chinese Opium Smokers (1880)
NoGAN training is crucial to getting the kind of stable and colorful images seen in this iteration of DeOldify. NoGAN training combines the benefits of GAN training (wonderful colorization) while eliminating the nasty side effects (like flickering objects in video). Believe it or not, video is rendered using isolated image generation without any sort of temporal modeling tacked on. The process performs 30-60 minutes of the GAN portion of "NoGAN" training, using 1% to 3% of imagenet data once. Then, as with still image colorization, we "DeOldify" individual frames before rebuilding the video.
In addition to improved video stability, there is an interesting thing going on here worth mentioning. It turns out the models I run, even different ones and with different training structures, keep arriving at more or less the same solution. That's even the case for the colorization of things you may think would be arbitrary and unknowable, like the color of clothing, cars, and even special effects (as seen in "Metropolis").
My best guess is that the models are learning some interesting rules about how to colorize based on subtle cues present in the black and white images that I certainly wouldn't expect to exist. This result leads to nicely deterministic and consistent results, and that means you don't have track model colorization decisions because they're not arbitrary. Additionally, they seem remarkably robust so that even in moving scenes the renders are very consistent.
Other ways to stabilize video add up as well. First, generally speaking rendering at a higher resolution (higher render_factor) will increase stability of colorization decisions. This stands to reason because the model has higher fidelity image information to work with and will have a greater chance of making the "right" decision consistently. Closely related to this is the use of resnet101 instead of resnet34 as the backbone of the generator- objects are detected more consistently and correctly with this. This is especially important for getting good, consistent skin rendering. It can be particularly visually jarring if you wind up with "zombie hands", for example.
Additionally, gaussian noise augmentation during training appears to help but at this point the conclusions as to just how much are bit more tenuous (I just haven't formally measured this yet). This is loosely based on work done in style transfer video, described here: https://medium.com/element-ai-research-lab/stabilizing-neural-style-transfer-for-video-62675e203e42.
Special thanks go to Rani Horev for his contributions in implementing this noise augmentation.
This is a new type of GAN training that I've developed to solve some key problems in the previous DeOldify model. It provides the benefits of GAN training while spending minimal time doing direct GAN training. Instead, most of the training time is spent pretraining the generator and critic separately with more straight-forward, fast and reliable conventional methods. A key insight here is that those more "conventional" methods generally get you most of the results you need, and that GANs can be used to close the gap on realism. During the very short amount of actual GAN training the generator not only gets the full realistic colorization capabilities that used to take days of progressively resized GAN training, but it also doesn't accrue nearly as much of the artifacts and other ugly baggage of GANs. In fact, you can pretty much eliminate glitches and artifacts almost entirely depending on your approach. As far as I know this is a new technique. And it's incredibly effective.
The steps are as follows: First train the generator in a conventional way by itself with just the feature loss. Next, generate images from that, and train the critic on distinguishing between those outputs and real images as a basic binary classifier. Finally, train the generator and critic together in a GAN setting (starting right at the target size of 192px in this case). Now for the weird part: All the useful GAN training here only takes place within a very small window of time. There's an inflection point where it appears the critic has transferred everything it can that is useful to the generator. Past this point, image quality oscillates between the best that you can get at the inflection point, or bad in a predictable way (orangish skin, overly red lips, etc). There appears to be no productive training after the inflection point. And this point lies within training on just 1% to 3% of the Imagenet Data! That amounts to about 30-60 minutes of training at 192px.
The hard part is finding this inflection point. So far, I've accomplished this by making a whole bunch of model save checkpoints (every 0.1% of data iterated on) and then just looking for the point where images look great before they go totally bonkers with orange skin (always the first thing to go). Additionally, generator rendering starts immediately getting glitchy and inconsistent at this point, which is no good particularly for video. What I'd really like to figure out is what the tell-tale sign of the inflection point is that can be easily automated as an early stopping point. Unfortunately, nothing definitive is jumping out at me yet. For one, it's happening in the middle of training loss decreasing- not when it flattens out, which would seem more reasonable on the surface.
Another key thing about NoGAN training is you can repeat pretraining the critic on generated images after the initial GAN training, then repeat the GAN training itself in the same fashion. This is how I was able to get extra colorful results with the "artistic" model. But this does come at a cost currently- the output of the generator becomes increasingly inconsistent and you have to experiment with render resolution (render_factor) to get the best result. But the renders are still glitch free and way more consistent than I was ever able to achieve with the original DeOldify model. You can do about five of these repeat cycles, give or take, before you get diminishing returns, as far as I can tell.
Keep in mind- I haven't been entirely rigorous in figuring out what all is going on in NoGAN- I'll save that for a paper. That means there's a good chance I'm wrong about something. But I think it's definitely worth putting out there now because I'm finding it very useful- it's solving basically much of my remaining problems I had in DeOldify.
This builds upon a technique developed in collaboration with Jeremy Howard and Sylvain Gugger for Fast.AI's Lesson 7 in version 3 of Practical Deep Learning for Coders Part I. The particular lesson notebook can be found here: https://github.com/fastai/course-v3/blob/master/nbs/dl1/lesson7-superres-gan.ipynb
There are now three models to choose from in DeOldify. Each of these has key strengths and weaknesses, and so have different use cases. Video is for video of course. But stable and artistic are both for images, and sometimes one will do images better than the other.
More details: