Hi Friends! đź‘‹
📊 Daily Stand-Up: Sorry for the long pause but rest assured we’ve been hard at work on our game. Last time, as the Lead Engineer, we decided that the game didn’t feel quite complete without images. So we built a system within the game to categorize and show images to the user along with a few images to test against. We also built a python script that created images for us. After our successful test, it was time to create the bulk of the images for our game. Our game now has over 1200 images at its disposal and we learned a lot about the capabilities of Dall-e v3.
Did you know when creating images of humans Dall-e has a bias towards young women? Well in our testing it did. It also has a bias towards beautiful, imo, people in general.
We spent $ 50+ to learn the following…
Time for some definitions:
Unusable in the context of AI generated images in this blog post simply means images that aren’t believable, i.e. the average person would know the image is not of a real human.
Fever Dream in the context of AI generated images in this blog post means the image wasn’t in line with the prompt or Dall-e added random elements rendering the image unusable.
Okay back to our findings:
Well it’s true, during the creation of 200, (a very small sample size), of young, (age 21 to 30), women 1.5 out of every 50 images were unusable.
This in contrast to middle aged, (age 30 – 50), men, 200, (a very small sample size), where 3.5 out of every 50 images were unusable.
Dall-e wants to create beautiful people. This sort of makes sense considering Open AI likely trained Dall-e on social media images where people tend to show themselves in the best settings, and publicly available images of celebrities etc
(Even when prompted to be sad they’re still super models)
(Reality lol)
Dall-e doesn’t factor in weight. In our original testing of the integration of images in our game we thought we would try weight as a factor in image creation of people. However Dall-e doesn’t seem to respect weight despite wording prompts with several instructions to respect all physical characteristics given.
Dall-e is weak when creating the eyes of humans. The human eye contains the iris, the pupil, the cornea, lighting bouncing off the cornea, etc. Dall-e often fails to create correct eye shapes, and more often than not correct light shapes on corneas ie the light reflecting off of the left eye should be nearly identical to the light reflecting off of the right eye.
(Left and Right eyes respectively with incorrect light reflections)
Dall-e is weak on human face outlines. Often we saw human faces with outlines that contained matrix-like glitches where images were simply not complete.
(Matrix like glitches)
4 in every 50 images required cropping. Despite asking for portrait style images with a single person, Dall-e often creates images with multiple faces, often the same face with small changes or different views of the same face.
Despite even the most specific prompting, Dall-e still has “Fever dreams”. We prompted Dall-e to give us images of realistic humans; However, 1.5 in every 50 images was unusable, i.e. an animated character, a person with 2 faces and other even more fantastical quarks. 2 mild examples:
(Animated human)
(Random plant in their pocket?)
Open AI has rate limits on Dall-e. In other words you can only create so many images per minute. These are based on spending and time with product tiers.
Conclusions:
All said and done we created 1200+ images, we kept 1200. We threw away close to 200. We projected our cost to be $ 50.00 USD and we almost hit that totaling out $ 55.50 USD.
So all in all Quality Control is required when generating Dall-e images at scale. And if you want help generating AI images for your product let’s talk because we learned a lot of lessons and can save you a lot of trouble.
Tomorrow we’ll talk about the integration of RevenueCat to support our in-app purchases.
💪🏼 New Skill: Dall-e prompt engineering
🎵 Playlist: https://open.spotify.com/track/36vTL8WP4ZoD1bbUaW0V3E?si=fa414c650dce42c5
🧠Mental State: Dall-e had more to teach me about AI than I thought. I’m excited!
🛠️ Tools: MacBook Pro M2, Ipad Pro, Pixel 6, Intellij, Flutter, The Keyboard that Saved my Hands, The Desk that Saved my Back, The Perfect Stand for Ipads, Desk Mounts that Work and Go on Sale, Ollama
đź“š Resources:
https://platform.openai.com/docs/api-reference/images/create
https://platform.openai.com/docs/guides/rate-limits/usage-tiers?context=tier-one
#AI, #TechJourney, #Flutter, #MobileGameDevelopment, #SoftwareDevelopment, #ProgrammingLife, #TechBlog, #AIGame, #AI Murders, #AI Kills, #AI Murder, #Defeat AI
By: David Giametta
Future David here! I hope you enjoyed reading through this post. AI Murders is now live! Apple, Google
Leave a Reply