I've been doing some (very) initial dabbling into AI generated images. Using various websites, I've been trying to generate accurate representations of classic cars. How come they do such a poor job?? If the car in question is very well known, a classic Mini for example, results can be great. However if a less known model (let's say a Hillman Imp) then it really struggles. What I don't understand is there is tens of thousands of pictures of Hillman Imps out there. Why can't the AI replicate these half decently? What am I missing?
Midjourney - AI art
- Thread starter tang0
- Start date
Soldato
Perhaps add in as many tags as you can for the Hillman Imp...
Last edited:
Soldato
Associate
I had the same issue with my wife's car (Seat Leon mk2). Could only get images of the mk3.
Caporegime
I love the bushes growing horizontally out of the hillside. Some of them are sprouting directly from rock. Also, why is the car empty? Who's driving?!
Last edited:
Soldato
Try playing around with the years in production and model names that were pushed at that time.
Soldato
MS Paint now has AI
There are way fewer images of that type of car than of a classic Mini in the training set so if you are seeing issues in the tool you're using it's likely down to that.
It doesn't seem too bad though, a quick test with DALL·E 3. It recalls the general details but seems to have hallucinated the grille and the style of the front a bit:
Those differences are just down to the lack of images of that car vs the mini.
If you add parameters in say a tool like Midjourney that's generally to impact the style of the image, it's not likely to have much impact on the ability to draw some object from memory if it's less familiar with it than some other object.
Last edited:
Caporegime
I've noticed a lot of these AI image generators now have guidelines in place that stops them from depicting certain public figures.
There might be a way around it, especially when using offline models but I had no such issues a few months ago.
There might be a way around it, especially when using offline models but I had no such issues a few months ago.
Soldato
That's probably for the best
Caporegime
I don't think Adobe Firefox really understood the assignment:
"A Hillman Imp driving on a mountain road"
"A Hillman Imp driving on a mountain road"
what is going on with that busThere are way fewer images of that type of car than of a classic Mini in the training set so if you are seeing issues in the tool you're using it's likely down to that.
It doesn't seem too bad though, a quick test with DALL·E 3. It recalls the general details but seems to have hallucinated the grille and the style of the front a bit:
Those differences are just down to the lack of images of that car vs the mini.
If you add parameters in say a tool like Midjourney that's generally to impact the style of the image, it's not likely to have much impact on the ability to draw some object from memory if it's less familiar with it than some other object.
Soldato
.
Last edited:
Caporegime
Now this i might try!
Associate
i've been playing around with Stable Diffusion, using the Easy Diffusion frontend. i have it locally so it runs on my GPU. it's an RTX 2060. so not the fastest, or card with the most VRAM card out there, but amongst the settings is the option to use minumum VRAM which, when using certain ControlNets is vital
i have this weird obsession with generating cats in hats (don't ask me why) but when i use text to image i've also tried other prompts. lightning has been another angle i've enjoyed
i have this weird obsession with generating cats in hats (don't ask me why) but when i use text to image i've also tried other prompts. lightning has been another angle i've enjoyed
Everyday the dead internet theory becomes more and more of a reality. Youtube shorts are already overrun with Ai voice overs and collages of AI images.
Soldato
Does anyone have a link for this site please or know what has happened to it? I used it a couple months ago to create some imagery and now I can't seem to find the same website again.