AI - Working with generated photos
Prompting - the act of providing cues to stimulate or encourage someone to take a specific action. At first glance this seems trivial, but a computer doesn't have the same assumptions as humans do. After many hours prompting with Midjourney and examining the generative art it created I have a couple observations I want to share with you.
Please notice the floating wine glasses and almost 3rd arm extending from the models body. This is the effect of my first prompt. The starting prompt was "a photo of a male model, in tuscany, drinking wine, golden hour"
All images seen in this post are AI generated
Precision vs Artistic freedom
When crafting prompts for AI, precision is key. A well-defined and specific prompt can help narrow down the desired response and increase the chances of obtaining accurate results. At least that was what I was told, turns out thats not quite true. So maybe a simple prompt?, give the reigns to the AI model?, let it figure it out? But here we risk misinterpreting our idea.
When I first heard that we'll see the rise of "prompt" professions, experts in communicating with AI, I blew a substantial amount of air out my nostrils in disbelief. Until I conducted 2 case studies with AI photography. There is a specific way you need to speak with computers.
I modified my initial prompt to include cypress trees, and I wanted to enforce the "drinking" part of the prompt. But as seen in this image, the AI went somewhat crazy ... The 4th image is completely off the scale.
The nuances of language
Ambiguity poses a significant challenge when prompting AI. Natural language is nuanced and often context-dependent, making it difficult to predict all possible interpretations of a prompt. AI systems may struggle to grasp the intended meaning, leading to unexpected responses.
I abandoned the verbs, AI struggles with that. Ideas like "pouring, holding, drinking" is too abstract for the current models. They know what things are, but not how they work. In this prompt remix, I lost the cypress trees which are a staple of Tuscany. In this variation my model is holding the wine glass correctly, and I got lucky with the fingers.
I lacked something interesting, appealing in the photo, so I decided to increase the amount of subject.
Refine, Refine, Refine
Effective prompting extends beyond technical considerations. User-friendliness and intuitiveness play a significant role in maximizing AI's usability and impact. Designing prompts that align with users' mental models, language proficiency, and expectations enhances their ability to interact with AI systems comfortably.
The Dalmatian is a nice touch, elegant, appealing ... perfect. I even got the trees in the background that I asked for. Animals, and especially cats are AI's favorite subjects to generate. AI models were trained what seems like mostly on animals and women :) When you ask for cat pictures or photoshoots of ladies you get near perfect results first time around. But when faced with something less popular, like for example a syringe it fails. No matter how I prompted it seamed Midjourney did not know what a syringe is.
How to get the most out of a prompt?
Be short, sweet and aware of where you place words.
For Midjourney, start with the type of image (photo, illustration, tattoo ...), follow it by your subject (a male, a dancer, a cat). Afterwards, what I've found is that you use single words to describe what else you wish to see in your image. End with styled effects, lighting, mood and all the other parameters you want.
"a photo, 70 year old, male model, in Tuscany, drinking wine, Dalmatian at his side, golden hour"
"Prompting AI is a nuanced art"
Prompting AI is a nuanced art that requires a careful balance between precision, openness, fairness, and adaptability. Crafting the right prompt demands a deep understanding of the AI system, context, and user needs. As AI continues to shape our interactions and decision-making processes, the ongoing refinement of prompt designs will be instrumental in unlocking the full potential of AI technology