Using AI to enable prompt-based avatar “remixes” in Slack
How do you help your company use and understand AI? Make it easy, make it fun, and make sure they can’t miss it. At Grow, Slack is our primary communication tool, where GIFs, emojis, and memes define our daily conversations. So we asked, “Could AI let us literally turn each other into part of those conversations?”
Everyone has a Slack profile image, so we started there, making a Slack service that allows you to “remix” a coworker’s image into whatever you want with generative AI. It was as simple as typing “/remix @neil as a superhero” when he was taking on extra work, turning Neil into something surprising and relevant with Stable Diffusion (and a little extra coding magic).
We used Pix2Pix (plus Stable Diffusion) to process the Slack requests, pulling the tagged user’s profile image from Slack and applying a prompt derived from the submitted Slack message. As with the current suite of generative AI, the results ranged from hilariously mind blowing to “hmmmm, let me try again in a smarter way.”
By obscuring the technical side of AI, we encouraged our team to work directly with AI for better results, all in the pursuit of a funny image. It was the perfect intersection of playful experimentation and intuitive learning. Teammates shared both the crazy “misses” (think horrific extra limbs and eyes) and the perfect interpretations of intent (/remix @kali as a cross-stitch).
In just 2 weeks, our team of 50 employees made more than 1,800 images, sharing the power of generative AI with each other, and learning both the power and the limitations of where the technology currently resides. For example, we found that the best results were when we asked to remix employees in a distinct style.
Asking AI to interpret distinct actions (surfing, weight lifting, coding, etc) often were disappointing – Stable Diffusion had trouble changing the poses/motion of static Slack avatar images. Also, Slack avatars that were less traditional (close-up facial images, side profiles, full-body images, pics with distinctive props/objects) were more problematic for creating exceptional results. We actually saw a large percentage of employees updating their Slack profiles with more classically framed headshots in order to have more versatile /remix results.
Overall, this experiment was a great success for empowered learning. Our team embraced both the possibilities and challenges of generative AI, showing themselves what could be the future when integrating AI into our everyday digital interactions.