I Gave the Same Prompts to ChatGPT and Midjourney… Who Won?

GPT understands better, but I don't think they will focus on improving image generation, at least on ChatGPT itself. Maybe they will launch a standalone image generator app with a lot of features, and that would definitely kill midjourney.

Damn, Midjourney has been training everyday with people generating stuff with their daily keywords. And it wasn't enough.

One thing that annoys me is when you inpaint to change one thing, ChatGPT regenerates (and frequently changes) the entire image. It should keep as much of the unpainted region as it already is and just adapt the region around the inpainted area. Hopefully this is on their todo list

The key is to combine them.

I’ve Been doing some experiments taking MJ prompts and and using ChatGPT to translate the prompts to work in other models and platforms. Using the same prompt (something I’ve tested for years now) doesn’t really work well. Results often are drastically different. Taking the Midjourney prompt along with a few sample results I have chatGPT also do an in depth analysis of the styles and techniques. I then have it create prompts for Reve, Leonardo Ai, NIM Ai and ChatGPT. The prompts are often very different since the tokens that each model understands is often very different. Results can differ depending on the subject. Photography styles work best doing this. I find most other platforms don’t handle illustrations as well as MJ / Niji , or at least in my tests to date. The idea of “ Hey learn this prompt” is meaningless since each platform understands language in a slightly different way which to be honest is annoying. Having to learn the quirks of each platform/ model is too much. But using ChatGPT as a translator helps and makes the process more accurate.

Nice video. Many videos deal with "consistent character", but a video testing "consistent background" in Midjourney, ChatGPT, Leonardo, and ComfyUI would be very interesting.

Human poses in Midjourney, try making the subjects facing away from the camera and... Sheesh.

if we could have chatgpt precision with midjourney style... that would be awesome

Also one good test is to try to make giants and giants interacting with normal size people. Such fantastical things MJ STILL fails at that after all this time. ChatGPT nails it easily

Something I've been trying is: literal visualization of "putting the cart before the horse", they get it wrong every time.