Creke told us that they've been busy these past few weeks improving Selfie 2 Waifu after work. “I divided the optimization work into 3 parts: refactoring architecture, improving computing performance, and reducing server error rate,” they said.
The programmer went into detail about their process: “In the first version, I put everything into one module just to make it simply work. As users increased, the error rate was increasing, too. Then I split the web frontend and computing backend as independent modules and made a distributed architecture. Then we could do the ‘more users, more machines’ trick.”
They continued: “Soon I realized that GPU servers are far too expensive, my credit card bills were flying like a rocket. So I gotta dive deep into the computing principle level, in order to save some money. As a result, a pre-fork multi-worker pipeline was built where GPU concentrates in model inferencing and CPU does the image decoding and resizing job.”
However, Creke admitted that they didn’t save money by doing this because every time the server performance got optimized, they got more visitors coming to their website, wanting to turn themselves into anime heroes. “But I'm glad I managed to host 3 or 4 times the audience by adding 1.5-2 times [the number] of machines, and the servers are running at a low error rate. Well most important of all, I'm so happy that you guys are enjoying it, that's all that matters.”
We also spoke to the programmer about privacy concerns. Here’s what they had to say: “For the algorithm, I am just using the pre-train model of UGATIT. The dataset of the official model is trained using 3,400 photos and 3,400 anime drawings from public dataset. Indeed, I want to finetune the model, 'cause now shadow makes some bad cases. But since I cannot collect any users' selfies from the site without their permission which results in short training data, the original model is still being used. So I have to optimize the post-process procedure of these cartoon drawings as much as I can.”
Creke said that they've set up a “basic but important principle about data and privacy” for their website. “We will not save any selfie without permission, all privacy-related actions will be clearly notified. Also, Google reCAPTCHA is used to protect and verify users' actions, which enhances security, and therefore Google's privacy policy applies.”
For the best results, it’s advised that you submit a passport-style photo with a plain background. While the website can turn anyone into an anime character, it works best with female selfies, just like the name of the website suggests. Keep in mind that you can get slightly different results if you upload the same photo twice, so if you decide to become a waifu, try it out a few times.
The online waifu tool is based on a 2019 paper by South Korea’s Clova AI Research, gaming company NCSOFT, and the Boeing Korea Engineering and Technology Center, titled "Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation." If it sounds complicated, it really is.
Some people also had fun by posting pictures of celebrities and pop culture characters instead of their own photos to see how the Selfie 2 Waifu tool responds. Some of the results are absolutely hilarious.
