This article aims to compare the reverse prompting capabilities of GPT-4V versus Midjourney. As a running example, we’ll use pics of a beautiful mansion in San Francisco. As a fun fact, sources like this and this say this is indeed Sam Altman’s house.
Altman’s Mansion
First, I passed this image into ChatGPT and asked it to create an image prompt.
ChatGPT suggested this prompt:
🧑💻 ChatGPT Image Prompt
Create an image of a luxurious multi-level house perched on a hillside during twilight. The house should have a traditional upper section with dark wooden siding and a modern lower section with expansive glass walls and concrete details. The interior lighting should be visible, casting a warm glow against the dusk sky. Lush greenery surrounds the house, including well-maintained bushes and mature trees. Include a set of sleek, concrete stairs with glass railings leading up to the entrance. The backdrop should feature a gradient of the evening sky transitioning from blue to hints of orange at the horizon, and the foreground should have a slight urban feel with the presence of distant city buildings.Feeding back the image prompt into it, ChatGPT creates the following image:
Looks similar already.
Now, passing the image into Midjourney, it comes up with these image prompts:
These are the 16 images (4×4), Midjourney generates when passing these image prompts back into Midjourney:
While all images are pretty beautiful, I don’t think they capture the essence of the real villa in the Lombard Street that’s more classy and less modern.
Passing the ChatGPT prompt into Midjourney also doesn’t help. It creates these images that are still pretty different:
Let’s try for the living area of Sam’s $27 million mansion next!
Altman’s Living Area
Passing this prompt into GPT-4V generates the following image:
Not so bad!
Next, I use the /describe command in Midjourney to create a prompt based on the given image. It creates the following prompt:
Wow, Midjourney figured out that San Franciso is the city in the background! This is amazing – DALL-E couldn’t do that.
Hitting “🥳 Imagine all” creates the following 4×4 image suggestions:
The images are beautiful but very different from the original. Let’s try the sauna next! 👇
Altman’s Sauna
Again, I’ll start with ChatGPT (GPT-4 Turbo):
ChatGPT recreates this image:
Even prettier than the original!
Let’s try Midjourney on Discord:
This gives us the following four prompts:
Hitting 🎉 Imagine all again:
Great inspirations! Remind me again: What do we need architects for when generative AI can create stunning designs like these? Just kidding. 😉
Check out the following learning path to stay on the right side of change and participate as a tech disruptor! 👇
🧑💻 Recommended: 10 Steps Learning Path: Becoming a Python Prompt Engineer 🚀 on the Finxter Academy