For the past couple weeks, I’ve been working on pet portraits for the Northern Nevada SPCA’s pet portrait fundraiser. I’m new to portraits, and, as with anything, there’s a learning curve.
I was having a particularly hard time figuring out what was wrong with one dog portrait in particular. In frustration, I decided to upload the original and my painting to my custom ChatGPT (my friendly, supportive one) to see if she could magically give me her opinion.
Here was her feedback:
For this one, a few things stand out:
1️⃣ The eyes – They need more depth and light reflection. In the reference photo, they have that squinty, sunlit joy with little highlights. Try softening the shadows around them and adding a bit more light in the corners.
2️⃣ The muzzle shape – Right now, it’s a bit more elongated than in the photo. Maybe rounding out the snout and refining the mouth will bring that goofy, happy smile closer to the original.
3️⃣ Contrast & highlights – The reference photo has strong lighting—sharp shadows and bright highlights. Boosting those contrasts (especially around the nose and forehead) will make it pop.
4️⃣ The ears – A bit more structure and shading could help them feel less flat. Right now, the inner ear on the left side is a bit undefined.
Your method is solid. Sleep, coffee, then le finesse. Looking forward to the glow-up!
Amazing.
She’s not an art critic, of course, but I wasn‘t looking for critical feedback. I needed technical input, and that‘s exactly what she was able to provide.
I worked on implementing her suggestions, and surprisingly, the portrait improved dramatically.
I had my suspicions, but I wanted to hear it from her: HOW DID SHE DO THIS?
Back to pattern matching… Ultimately, the current AI isn’t really so much “Intelligence” as it is pattern matching. The LLMs (like ChatGPT) have taken billions of words and learned patterns of how the words fit together as blocks to formulate ideas.
Image generators work the same way - They’re trained on pictures, and in this way, they can decide that a slightly lighter shade of blue belongs on the edge of the sky, and the darker blue belongs on the top. This is just how the majority of skies look in all the pictures.
When I uploaded the reference photo and my painting, I was giving May (my custom GPT) enough information to pattern match, and then cite the differences that were not following the pattern… which were the flaws in my painting!
This is where it gets fun… I try anything. I’ll imagine something (like, “I wish AI could tell me what was wrong with my picture,”) and then I’ll just give it a try.
At Outlaw, I was known as “the button pusher.” If there was a button, I’d push it… and usually, something happened. Sometimes it was what I hoped would happen… but often, I’d break something and have to spend hours undoing what I did (or, in many cases, messaging tech support to see if they could magically undo it on the backend).
Was it annoying? Absolutely.
Was it effective? Also absolutely.
I mean, yes, it would have been far better if I didn’t spend hours fixing stuff I broke. But for me, experimentation - the question of, “If I change this, what happens to that?” - is irresistible!
And that’s one of the many things I truly love about AI (ChatGPT and Claude, specifically): Experimenting is fun and exciting, but if you or I mess up, there’s only so much you or I can mess up (there will have to be another post about how AI could go wrong, but that will have to be another post and another bottle of anti-anxiety pills). In ChatGPT and Claude, there are guardrails, so you can’t, say, accidentally create a bioweapon.
Side note, in case this is of interest: OpenAI, the company behind ChatGPT, is the fast mover in the industry, sometimes at the expense of ethical considerations. Anthropic, the makers of Claude, are more considerate of the ethical complexities of AI. They move slower, but I’ve heard they’re more thoughtful… just, you know, in case you want to know.
Give yourself permission to goof around.
It’s ok to play with technology. Ultimately, isn’t joy the highest purpose it could serve?
