Plastic or Fantastic: Achieving Emotional Expressions with AI

James Dow

Rethinking Advertising: Emotions

The ability to convey authentic emotions is a cornerstone of compelling storytelling. Faces tell stories, and expressing nuanced emotions in portraits is crucial to eliciting a genuine connection with your audience. 

Before the advent of Gen AI, emotional campaigns required talented models and intuitive photographers. That’s no longer the case. Capturing an array of emotions on an array of models is now achievable in minutes. How, though, does AI stack up to the real deal? 

The Creative Challenge

There’s a prevailing belief that’s snaked its way through adland: AI can’t do emotions. A lot of that skepticism stems from AI's capacity - or incapacity - to authentically portray facial expressions. Rather than realistic portraits, we’ve seen homogenous faces that feel plastic or unrealistic. It’s been referred to as “AI face”, in which you get a semi-pleased model with an eerily artificial expression.

We wanted to explore Pencil’s capabilities when it came to conveying a wide spectrum of emotions on the same model. The test was about showing emotions, but also about ensuring consistency in the character depicted across varied expressions. We wanted to move beyond superficial emotions and create genuine, relatable faces that would capture the attention of audiences.

The Outcome

Getting a model or character to smile using AI is simple enough - but we wanted to see if we could capture personality in their face. In the example here, you get a sense of the cheeky personality that accompanies the smile. The creases around the smile give an impression of spontaneity, along with the ruffled hair and lean towards the camera. You can almost hear the model’s gruff laugh. But the shot isn’t overworked. You still get a sense of realism with the model’s stubble, and his speckled skin. It’s a far cry from the overly smooth, plastic skin historically associated with AI. He’s less Barbie’s Ken, and more ‘bloke you knew from school who’s had a glow up’. There’s realism and texture to the image.

Importantly, though, we were able to see other emotions from the same model, including a more sincere look. There are a number of brands that trend towards serious imagery, and you can see the benefits of that style in the image below. As an image, it’s a lot more earnest. The way you’re drawn in is different. It’s more about inspiring empathy or curiosity, with a less clearly-definable expression, and a lot more nuance to the emotion being portrayed. 

We also wanted to push our model into less common territories. The anger or despair in these examples shows that you can use AI to translate complex emotions, using the same model. In the same way that you could almost hear the model’s laugh, you can now hear him yelling. It’s the level of detail - the creases, the furrowed brow, the gaping jaw - that give the image depth.

The Methodology

Made by @JamesD0w (Creative Director, Brandtech Consulting). Here’s how he did it:

1. Prompts

  • A sports photo of a young male athlete smiling in a bright pastel pink backdrop with dynamic angled lighting and neon blue. contrast. 
  • A sports photo of a young male athlete shouting in a bright pastel pink backdrop with dynamic angled lighting and neon blue. contrast. 

2. Images

Let’s create the first character we will build from. I’m doing that in DreamStudio today, using Stable Diffusion XL. Let's break the prompts down:

  • ‘Smiling’ will be the word we change each time. Smiling is nice and simple to begin with.
  • Adding ‘dynamic angled lighting’ helps create a dramatic aesthetic. 
  • Choosing ‘Neon blue’ helps with the sports energy.

OK so we have a smiling athlete. Let’s change his mood. Maybe he lost the game? Let’s swap out the word ‘smiling’ for ‘angry’ in the prompt. 

We also need to make a variation of this image so we set this as our initial image. This means the model will base our new prompt on this existing image but use the new word ‘angry’ to change it.

So that’s the process, play around with different emotions and don’t be afraid to hot ‘Dream’ a lot! Getting a good image can take between 20 to 40 images sometimes. Wait for the right one!

3. Upscaling

Finally we upscale all the preferred images using Magnific AI on the following settings:

  • Creativity 1
  • HD 5
  • Resembalnce -1

The Conclusion

The tests aimed to challenge skepticism around AI's capacity to authentically represent nuanced expressions, often criticised as rendering faces with a plastic or artificial quality. Our test sought to transcend these limitations by ensuring not only the conveyance of genuine emotions but also the consistency of character across varied expressions. In the same model, we saw fury as well as joy - and we felt the personality behind both images. Gen AI is continuing to bridge the gap between technical detail and authentic emotion, offering a tool that not only captures a diverse array of feelings but also ensures that these emotions are conveyed with depth, subtlety, and authenticity.