After that you'll be shown with two windows, one showing the original image and the second showing the colorized image, you just have to press any key to close both windows but remember to right click on the image if you wish to save it.
WARNING: Video files can be very heavy to process But if you want to give it a go then you should just run: python bw2color_video.py \ –prototxt model/colorization_deploy_v2.prototxt \ –model model/colorization_release_v2.caffemodel \ –points model/pts_in_hull.npy \ –input images/your_video.webm
This one should give you a look at the colorized frames, those frames will be stored in a folder called frames and you can join them to make a video with the "make_video.py" file, just remember to adjust the framerate.
Some of these results are obviously shitty. You could crop parts that didn't work properly and try that piece seperately.
I already tried that with many images but the result was either the same or even worse, the thing is, this neural net needs a lot more training samples and time, I'm pretty much testing it on every image I can find and seeing what happens.
Can we retrain the model? Use the one they have as a base but feed it a bunch of moden re-enactment photos.
Yes, you can do that, all you gotta do is call all the components of the model that were saved in the last checkpoint (the files in the model folder) and run a training session with new images to feed the network. The only problem is that I don't know yet how to do that with caffe models, gonna have to study that.
Does it work like deepfakes? Used to have that pumping for a while.
Close to it, this model uses a simple Encoder -> Decoder architecture, which honestly doesn't give too good results, however it is more faster and less expensive than what deepfakes use.
Deepfakes use GANs (Generative Adversarial Networks) to create their output, in short, it consist of one network creating fake data and another network inspecting said data and determining if it's real or fake, you basically have a network training hard to jew the other one to output a fake image that looks legit to us. This the kind of network I'd recommend for the colorization problem, it would give better results but as you've might imagined it requires a ton of GPU power and millions of samples.
This is really cool OP, thanks for sharing. It seems blurry spots and skies don't always turn out well but overall it is very impressive technology. Posting some of the better ones.
@echo off set /P id=Enter image name: python bw2color_image.py –prototxt model/colorization_deploy_v2.prototxt –model model/colorization_release_v2.caffemodel –image images/%id% –points model/pts_in_hull.npy pause null
in the folder containing the bwtocolor_image.py file and run the batch file. Then run the batch file and type in the name* of the image you want to convert. *include the .jpg, .png, etc at the end and make sure the image is in the images folder.
>File "bw2color_image.py", line 5, in Thanks anyway user but i'm still too retarded. I've gone back and looked through the command and everything and don't know what the fuck this "numph" thing is. I found a site that does it pretty good online though. see pic related.
on the second to last line of the file. 3) Use the script I previously posed and watch as all of your colorized images are saved in the color_images folder without the snipping tool.
Have Fun & Heil Hitler
Didn't work bro. I didn't have PIP so i got a file online "get-pip.py" and installed that with python. it came up that it had been added and everything but i still get the same error when running your script from earlier about "numpy". not sure what i'm going wrong.
What version of python are you using? The most stable from what I remember is currently 3.5.2
3.7.3 I will try out with 3.5.2 but then I'm giving up (kek) appreciate your help though user and not laughing at me and calling me a faggot for not understanding
its not in your PATH system variable. you'll need to add the pip executable's path to your system's path so your cmd knows where to look for it. quick google search for "add python to system path" will work for you.
For example. You didn't get Hitler's sleeve. So I lassoed his sleeve and colorized it. Hue 243 Saturation 9 Lightness -48. You do this for each individual color. It's time consuming of course. Not automated.
No shit bro. I tried working on a image yesterday and was doing it color by color. It's going to take time if you want it done properly (pic related was a half hour or so). This algorithm is certainly not perfect; it does make the images look better though. It's not supposed to be perfect, yet, until the AI gets better anyway. For now though, this gets the job done quick and makes a vast improvement on the initial b&w image.
Images in black and white usually feel 'disconnected' from people living today in a sense. If we took images of Hitler's Germany and gave them color it would actually go far in humanizing a nation and a man who has been dehumanized by history. see
Appreciate the pics very much.
This is great
This is true, and I initially assumed it was the point of this thread.