ADP: Always Designing for People
Year: 2019
Agency: Havas New York
Project: AI-Generated Images & Videos
In today’s ever-shifting landscape of work, we re-branded ADP into an identity that strategically positioned it as Always Designing for People. We created a flexible design system that uses AI as a signature element to communicate ADP as an influential HR leader.
Using neural networks, I created their images and videos in collaboration with Machine Learning. I asked the machine: “How can you be the artist for ADP?”
// Exhibited: l
* SXSW 2019
Style Transfer: The Artist Behind ADP
ADP’s visuals are generated using a machine learning technique called Style Transfer - the technique of recomposing images in the style of other images. Style Transfer takes two inputs—a content image and a style reference image—and blends them using convolutional neural networks. To illustrate in this example, a content image is the ocean video, and the style reference image is the graphic on the bottom left. Style Transfer blends those images to create this final generated output on the right.
Images
Using Style Transfer, we came up with the Uplifting style as our style reference image. It uses tones and organic shapes referenced from ADP's new visual identity and comes in 3 colour schemes: Purple, Rose, and Tan. Training a computer on each input took approx. 8-9 hrs
While training takes hours, applying the style on an image takes merely 2 seconds. This was highly important because we needed to use tools that were scalable in order to produce content for such a wide-reaching project.
Videos
Our Style Transfer technique is built upon Manuel Ruder’s fast-artistic-videos method. It offered an implementation of video style transfer with improved frame-to-frame stability by adding optical flow*. This resulted in generating consistent and stable stylized video sequences even in cases with large motion and strong occlusion.
*Optical flow refers to the visible motion of an object and the apparent 'flow' of pixels in an image.
Stylizing videos was a complex process due to optical flow computation. Training one input took 16 hrs and stylizing a 30 second video took approx. 4 hours
Uplifting fits perfectly within the overall design guidelines for ADP. It is never used by itself. Stylized images and videos can be composed of a mixture between our shapes, our icons library, and our custom typography - giving it overall a polished look.
It is so much more than a simple visual effect. It effectively acts as a metaphor for ADP’s new campaign “What are you working for?” and as a symbol for who they are overall - how they use AI in their products and how they are focused on supporting systems that allow people to work.
Early Experiments
Before Uplifting, we experimented with many different models and styles. One technique we tested was built upon Hang Zhang’s Multi-style Generative Network for Real-time Transfer method. Sure we could’ve trained the system on a million images BUT when you take into account training on one image takes several hours - our experiments needed to be well designed and focused.
The Redprint Style: One of the first explorations was testing very fine patterns composed of lines and grids with variations in size and thickness.
The Iconic Style: This test was trained on a set of early ADP icons. They are composed of organic lines and separated blocks of colour. It replicates a more minimalist design approach and a broad use of white space.
The Kent Style: Continuing along the previous path, we tested organic figures and sets of colors, but this time used larger/bolder shapes and got rid of our strokes. The result was not as refined as the previous attempt..
The Uplifting Style (Early Version): We found that the variations in tones and overlap of organic and delicate geometric shapes yielded the most beautiful results.These were our very first passes with Uplifting. This style continued to improve and morph into the final version that we see today.
