DOCTOR
THIS MONTH THE DOCTOR TACKLES...
> Train AI with
> Malfunctioning me
> Encrypt single files
AI image generation
I’ve started using Adobe’s Firefly AI image generation system (https://firefly. adobe.com) to produce photo-realistic images to my specifications. However, despite my best efforts, I frequently struggle to provide a prompt that will deliver an image suitable for my needs. What tips can you offer for improving the output that Firefly produces?
—Marcus Lyles
THE DOCTOR RESPONDS:
While AI image generation is still relatively new, it’s capable of some stunning results —except when you start out using it, when you find yourself never quite able to capture what you’re looking for. It reminds the Doc of the early days of AI-powered image colorization, when early training models relied on humans to evaluate their work. Because the engineers involved couldn’t properly articulate their feedback, these models kept favoring ‘safe’ colors like brown over more realistic choices when generating their images.
The breakthrough occurred when a new model was developed—DeOldify—that delegated the role of both generator and evaluator to the computer, effectively handing over the training of both creator and critic to the other AI tool. It wasn’t long before the critic’s more precise observations delivered improvements to the images being generated, resulting in the generator producing more realistic colors using a wider range of hues.
How does this relate back to your problem? Simple: rather than articulate what you need yourself, delegate the role to another AI model. The Doc has enjoyed great success using ChatGPT (https://chatgpt.com) to help him write prompts to deliver the images he’s looking for. Start by describing what type of image you want in ChatGPT, and ask it to create a prompt for Firefly.