Does this sound like just another hymn for the endeared prisma?
It is, but it comes from a person who waited a month to dive into the prisma trend despite knowing its specialty.
It is, but it comes from a person who waited a month to dive into the prisma trend despite knowing its specialty.
Published on: August 08, 2016
One fine morning when I logged in to check my Instagram, I immediately noticed a selfie with an artistic filter with a watermark sitting in a corner titled Prisma. While I tried to scroll down considering it a mere filter or camera app just like Retrica surfaced years ago, something caught my eyes. Why is her hair strands not washed out like the other filters? Why is the brush strokes confined to her faceline as if it was carefully painted? With the questions bugging me, it didn't take much time to find yet another fine crafted selfie with Prisma watermark that's awfully similar to a painter's masterpiece. I knew at that moment, a new hype is emerging and it's here to stay.
Origin of Prisma
While Prisma is undoubtedly the first service to reach to the mass, it’s certainly not the first of its kind. Similar services have existed long before Prisma came around. One such service is DeepArt.io which was published on October 25th, 2015 [1]. The algorithm they used was developed by three researchers from the Bethge lab at University of Tübingen in Germany [2] which itself used the Very Deep Convolutional Networks for Large-Scale Visual Recognition developed by Visual Geometry Group (VGG) in University of Oxford [3]. Prisma co-creator Alexey Moiseenkov, who developed Prisma as a side project when he was working for mail.ru, openly confessed to be inspired from existing services [4]. Prisma is based upon the same research paper [5] as DeepArt along with another by Li and Wand [6] [7]. Moiseenkov’s innovation was to dramatically improve the way it works and present the service to the mass in the simplest form.
Other similar services are Dreamscope and Neural Style Online, both of which are based on Google’s DeepDream.
Other similar services are Dreamscope and Neural Style Online, both of which are based on Google’s DeepDream.
Why Prisma is the one to be successful
There must be a reason why we’re affected by Prisma-fever and not DeepArt-fever, and the key reason is usability. You see, DeepArt takes a minimum of 10 minutes to render the image in 500x500 px resolution in free version and about 24 hours if you want the 9MP version for 149€ [8]. But even though Prisma is based on the same principles, it took them two months to enhance the algorithm [9] to deliver the image in less than 5 seconds. And that’s huge, considering Prisma delivers a higher resolution image too (1080x1080 px). Prisma uses three individual neural networks [10], all of which performs separate tasks.
Then comes the fact that it’s available as a smartphone app. You know what else is a smartphone app? Instagram. And basically every social media has a smartphone app too. Just recently Instagram announced that they’ve reached 500million users, of which 300 million are daily active users [11]. Facebook has over 1.13 Billion daily active users, 1.03 Billion of whom use the mobile service too [12]. So what did Prisma do (and DeepArt didn’t) to target this huge user-base? They put the Instagram and Facebook sharing icon just below the rendered image along with save to gallery and sharing via other apps option.
While sharing the image, you have no obligation to keep the Prisma watermark or the #Prisma hashtag unlike DeepArt. But guess what. People didn’t want to lose the chance to show off. The free advertising army has posted over 2 million Prismafied images under the hashtag since its launch. The app itself is very minimalistic too, you must try really hard to mess something up.
Then comes the fact that it’s available as a smartphone app. You know what else is a smartphone app? Instagram. And basically every social media has a smartphone app too. Just recently Instagram announced that they’ve reached 500million users, of which 300 million are daily active users [11]. Facebook has over 1.13 Billion daily active users, 1.03 Billion of whom use the mobile service too [12]. So what did Prisma do (and DeepArt didn’t) to target this huge user-base? They put the Instagram and Facebook sharing icon just below the rendered image along with save to gallery and sharing via other apps option.
While sharing the image, you have no obligation to keep the Prisma watermark or the #Prisma hashtag unlike DeepArt. But guess what. People didn’t want to lose the chance to show off. The free advertising army has posted over 2 million Prismafied images under the hashtag since its launch. The app itself is very minimalistic too, you must try really hard to mess something up.
Image that made me fall in love with Prisma
Being an art lover and tech enthusiast, I found it really fascinating to see the images Prisma and DeepArt can create. But one single image made me fall in DeepLove with DeepArt and Prisma. I was browsing some photos I took on my visit to Sonargaon. Seeing the sculpture “The Struggle” inspired by the painting of same name by Legendary painter of Bangladesh, Zainul Abedin. Suddenly I thought how’d it look in painting? I immediately installed Prisma and tried a few presets. I was astonished. See the results yourself.
But then I’d thought how much it really looked like the original painting? That’s where DeepArt, Dreamscope and Neural Style Online come in with their option to manually input style image.
শিল্পাচার্য জয়নুল আবেদীনের তৈলচিত্র 'সংগ্রাম' (Sangram) "The Struggle" by Zainul Abedin
I uploaded the photo I took as the base image and the painting above as the style image. I couldn't believe the results I received! And yes, these three images are the reason I had to research on Prisma for two days and finally write this huge post.
I tried input the original image and a prismafied image as style in DeepArt. This is what I got as the result.
Although it messed up when I gave it the original painting by Zainul Abedin and the photo I took as the style.
Although it messed up when I gave it the original painting by Zainul Abedin and the photo I took as the style.
My Prisma love didn't stop there! While I found Deepart and Dreamscope to be better, like I said, Prisma became the successful service for a reason. I played around with some more photos.
The uniqueness lies in its intelligence. Artificial Intelligence.
My friend Suhail Najeeb’s post on Prisma was really intriguing. He took the time to explain how Prisma’s AI works, so I’d recommend reading his blog. Here’s an excerpt.
"Prisma itself is not a photo filter. Its an AI. Yes, you're hearing me right, its based on an Artificial Intelligence, as claimed in their official website. |
Prisma uses convolutional neural networks to recreate the entire image from scratch using the structure of the uploaded image and style of a preset. ConvNets try to mimic biological processes and have been used for long in facial recognitions [13]. While there are already thousands of AIs at work, what separates Prisma from others is now you can visualize the effects of an AI right in front of your eyes. And it is now painting arts which are comparable to creative arts by humans. DeepArt is closing on to passing the Visual Turing Test [14]. I took the quiz and failed to identify two images out of ten.
I found it interesting how both the researchers of VGG are now working for Google DeepMind after Google acquired their startup Vision Factory [15]. They are working to help Google improve its image processing system, which include object recognition in search, its camera-based search apps and the data-processing systems needed for its self-driving cars [16]. Remember how you freaked out when Google introduced object recognition in Google Photos and could detect your dog perfectly? These guys are behind it!
DeepMind also developed the AlphaGO [17] which beat the best human player, Lee Sedol, in the board game Go which baffled the tech community. It all works on machine learning. AlphaGo was fed thousands of Go game strategies. While the success rate was initially in single digit, later it became 99.8% [18]. The more examples it gets, the better it becomes. That's how machine learning works. That's why Google Photos can detect objects; it was fed thousands of dog images and was told, 'it's a dog'. And who fed the images to Google Photos? You! When you clicked on an image on NoCaptcha Recaptcha (developed by Google) on hundreds of websites, you are presented with images to choose an object from. It doesn't detect if you're a human by the images you choose, it (possibly) tracks your cursor movement, IP address, browser cookies, time intervals etc to ensure you're not a bot [19]. It also feeds the same sets of images to hundred other users to match if they all select the same images and when they do Google learns a new object! [20] You are literally contributing in developing Google's AI.
I found it interesting how both the researchers of VGG are now working for Google DeepMind after Google acquired their startup Vision Factory [15]. They are working to help Google improve its image processing system, which include object recognition in search, its camera-based search apps and the data-processing systems needed for its self-driving cars [16]. Remember how you freaked out when Google introduced object recognition in Google Photos and could detect your dog perfectly? These guys are behind it!
DeepMind also developed the AlphaGO [17] which beat the best human player, Lee Sedol, in the board game Go which baffled the tech community. It all works on machine learning. AlphaGo was fed thousands of Go game strategies. While the success rate was initially in single digit, later it became 99.8% [18]. The more examples it gets, the better it becomes. That's how machine learning works. That's why Google Photos can detect objects; it was fed thousands of dog images and was told, 'it's a dog'. And who fed the images to Google Photos? You! When you clicked on an image on NoCaptcha Recaptcha (developed by Google) on hundreds of websites, you are presented with images to choose an object from. It doesn't detect if you're a human by the images you choose, it (possibly) tracks your cursor movement, IP address, browser cookies, time intervals etc to ensure you're not a bot [19]. It also feeds the same sets of images to hundred other users to match if they all select the same images and when they do Google learns a new object! [20] You are literally contributing in developing Google's AI.
When I started writing the post, I really had no intention to dig this deep. But the more I knew from my research, the more I thought people should know these. I don't know how mass people feel about Prisma or AI, but I'm totally engulfed by the idea that we're spectating something which had been predicted for decades and is now formulating right before our eyes. This is the revolution I was waiting to see. While the Artificial General Intelligence is yet to come, when it gets here, it would either help us become the supreme power in the cosmos or destroy us for the greater good of cosmos.
References
- DeepArt Publish Date: http://blog.deepart.io/2016/01/13/hello-world/
- DeepArt researchers: https://deepart.io/page/about/
- 19-layer VGG network by Karen Simonyan and Andrew Zisserman
- Inspired by other services: https://themoscowtimes.com/articles/why-russias-prisma-app-could-take-over-the-world-53516
- Paper by Gatys, Ecker and Bethge: http://arxiv.org/pdf/1508.06576v2.pdf
- Paper by Li and Wand: http://arxiv.org/pdf/1601.04589v1.pdf
- Bloomberg: http://www.bloomberg.com/view/articles/2016-06-20/new-masters-software-has-painterly-technique-down-pat
- DeepArt Pricing: https://deepart.io/pricing/
- Took two months: https://techcrunch.com/2016/06/24/prisma-uses-ai-to-turn-your-photos-into-graphic-novel-fodder-double-quick/
- Three neural networks: https://techcrunch.com/2016/06/24/prisma-uses-ai-to-turn-your-photos-into-graphic-novel-fodder-double-quick/
- Instagram statistics: http://blog.instagram.com/post/146255204757/160621-news
- https://investor.fb.com/investor-news/press-release-details/2016/Facebook-Reports-Second-Quarter-2016-Results/default.aspx
- Convolutional Neural Networks: https://en.wikipedia.org/wiki/Convolutional_neural_network
- Visual turing test: http://blog.deepart.io/2016/02/10/visual-turing-test/
- Acquire Vision Factory: https://techcrunch.com/2014/10/23/googles-deepmind-acqui-hires-two-ai-teams-in-the-uk-partners-with-oxford/
- Help in image processing: https://www.theguardian.com/technology/2014/oct/23/google-uk-artificial-intelligence-startups-machine-learning-dark-blue-labs-vision-factory/
- Alpha-Go: https://deepmind.com/alpha-go
- Alpha-Go success rate: http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html
- How NoCaptcha ReCaptcha works: http://stackoverflow.com/questions/27286232/how-does-new-google-recaptcha-work
- http://security.stackexchange.com/questions/78807/how-does-googles-no-captcha-recaptcha-work
