TYMBARK: Campaign with deepfake technology
Tymbark has been accompanying Poles for decades. It is at breakfast, lunch and dinner. At the picnic, match and canoe. During birthdays, name days and weddings. Irreplaceable, beloved, ours. A true lovemark, full of positive emotions. How to express it? How to enhance the feeling? How to involve new recipients in it?
"Everything tells me that someone loved me". - With the words of the well-known song, Tymbark decided to express the feeling that unites it with the Poles. The Skaldow hit in new arrangements was recorded for the brand by Sanah, Artur Rojek, Vito Bambino and Kwiat Jabłoni. Thanks to the application we built, each user could decide which artist would sing the next verse of the song, as well as join the singers and perform with them on Tymbark's virtual stage. So how many users, so many versions of the music video.
In the first phase, we faced the key challenge of creating a deepfake technology solution that would generate personalized video in the shortest possible time, with the highest possible image quality. Quality, stabilization and speed of video creation was our priority from the beginning. The interactive video was to become a showcase for the campaign, so we counted on massive user engagement.
In the first phase of dev work, we analyzed dozens of off-the-shelf solutions available on the market, but none gave satisfactory results. During the trials, it turned out that they require gigantic computing power, and it takes im.... several hours to generate an animation with a user's photo. This was unacceptable. So we made the decision to build the mechanism from scratch.
From the beginning, we experimented in search of an optimal solution. First, we created a mechanism to perform machine learning calculations on graphics cards, and although it worked really well, it required the purchase of server space for several hundred thousand zlotys.
After many optimizations and the use of innovative caching of the layout of characteristic facial points, we achieved incredible results! We reduced the video clip generation time from the initial 50 seconds to just 2.5 seconds. We were able to do this despite the limitations of the server infrastructure. Although it supported a load of several hundred thousand users, the cost of using it dropped tenfold!
After three months of experimentation and testing, we were able to create software that took less than 2.5 seconds to generate a personalized, high-quality video. However, we recognized that even that much might be too long for some. In order not to lose engagement at the last stage, we used a UX trick familiar from airport queue management. Our system started generating a video as soon as a photo was taken. The unsuspecting user was then prompted to enter his name and choose the order of the artists. When he did so, we could immediately present him with a video using his photo.
We took care of the user experience from the very beginning of the work. We analyzed possible navigation paths and UX architecture. We aimed to make the use of advanced technology as easy and pleasant as possible. We reduced potential barriers to a minimum, such as the number of form fields needed to be filled out, and most importantly, we simplified the steps for generating and sharing video on social media.
visits to the wszystkomimowi.pl campaign website
X - number of personalized vidoclips
X - number of users who used deepfake technology in the application
Silver sword in Craft Audio category
Silver in the category Innovation in Communications - Tech
Silver in the category Innovation in Communication - Content/Entertainment
Award in the Digital category - Integrated digital campaign
Distinction in the Campaign category