Designing Cars With AI - Fun for a rainy day

[funkyscooter]

I have not spent much time with AI. It's amazing and totally frustrating all at the same time. Prompt ChatGPT with the same detailed description several times, and the results are often wildly different. But this video on YouTube got me curious. If you don't follow Wray's YouTube channel, its worth a look. He seems very excited with the possibilities of creating designs for custom cars with AI, and one of his team goes through a few approaches.
https://www.youtube.com/watch?v=PG3VM0JxMlE
3 different methods are described, from photo realistic to cartoonish. This is more at the rendered drawing end of things and it is the easiest to get started.
Go to vizcom.ai and create a free account. No sign ups. I just used my Google account details.
Process is simple enough. You can just add prompts and see what it creates. But giving it a guide to work off sure helps, and if you want to just do something like change the colour of a car then this is the way to go.
Upload a picture of a car and start adding prompts, then vary the drawing influence. %5 is mainly the image generated by your prompts, %100 is the generated image's attributes applied to your uploaded image. (I think). Each time you hit generate, even with the same prompts, it feels like it changes the result, so if you see something you like its best to save before continuing. You can select areas of the image to influence but for this test I just let it do its magic to everything.
I gave it this picture. Its nice and clean, and free of background influences which should make the AI's job a little easier. I did test this on other images with people in the background and it does a pretty good job of removing them most of the time.

Screen Shot 2023-11-24 at 1.04.00 pm.jpeg (163.95 KiB) Viewed 647 times

Mix the drawing influence down to %5 and it reveals what it generated from the prompt "1957 holden car". Why 1957?, well when I tried 1960 the cars generated did not have wrap around windscreens. Figured I would go with what it might know - '57 Chevy.

Screen Shot 2023-11-24 at 1.05.47 pm.jpeg (111.45 KiB) Viewed 647 times

Now at %60 percent mix you can see the influence of a generated car on the image. I removed holden from the prompt by this stage and added convertible, so the original image would have most likely changed, but you get the idea.

Screen Shot 2023-11-24 at 1.10.17 pm.jpeg (141.86 KiB) Viewed 647 times

Changed the colour to light green

Screen Shot 2023-11-24 at 1.10.58 pm.jpeg (133.22 KiB) Viewed 647 times

Set drawing influence to 100%, still a convertible, but tones down some of the other lines (Chev bump on the rear door).

Screen Shot 2023-11-24 at 1.11.35 pm.jpeg (137.47 KiB) Viewed 647 times

Then tried to change the rag top to black, could not get it to work so in frustration I imagined the car as a 1980 model. Yikes!

Screen Shot 2023-11-24 at 1.15.32 pm.jpeg (132.96 KiB) Viewed 647 times

Tried to Refine the image and give it wooden side panels but clearly my prompt skills are lacking - car colour also changed. Lucky I saved an image of the original 1980 inspired Holden, otherwise it would have been lost forever This one would not have looked out of place on the showroom floor next to a Lada Niva

Screen Shot 2023-11-24 at 1.21.18 pm.jpeg (138.82 KiB) Viewed 647 times

Mixing it back to %5, you can seen the AI generated inspiration a little clearer

Screen Shot 2023-11-24 at 1.21.52 pm.jpeg (139.1 KiB) Viewed 647 times

Not sure if this kinda thing inspires others but I just wasted 3 hours messing with this and could easily waste another 3.

[avm]

Thank you for sharing this very interesting piece about your AI experiment. I wonder whether the AI could be pushed further, for instance to recommend technical upgrades to meet stated specifications. For instance: " Redesign the Holden EK, including its chassis, to include the following engineered and ADR compliant upgrades: Seatbelts, collapsible steering column, disk brakes with ABS on all four wheels, engine power increased on 57 kW to 280 kW".

Do you think that may be possible with one of those AI tools?

[Harv]

Thank you for sharing this very interesting piece about your AI experiment. I wonder whether the AI could be pushed further, for instance to recommend technical upgrades to meet stated specifications. For instance: " Redesign the Holden EK, including its chassis, to include the following engineered and ADR compliant upgrades: Seatbelts, collapsible steering column, disk brakes with ABS on all four wheels, engine power increased on 57 kW to 280 kW".

Do you think that may be possible with one of those AI tools?

Yes, you can do that. I used AI (Andrew's Intelligence) and those requirements, and got this:

1962 Holden wagon.jpg (217.64 KiB) Viewed 639 times

Other than the missing ABS, and the 280kW sneaking up to 304kW, I reckon the AI did a fair job

Cheers,
Harv

[funkyscooter]

From my (very limited) experience, AI (well ChatGPT) is a bit like asking questions to an 8 year old with all the knowledge of the world, and absolutely no understanding of any of it. Start by asking a simple question, get a broad answer. Then refine, then refine, then refine and out of left field the results will start to come back totally changed. It can end up being a frustrating cycle.

vizcom.ai looks like it could have the potential to help visualise your ideas, but it's not there yet. Type EK Holden and you get this

Screen Shot 2023-11-24 at 3.52.20 pm.jpg (95.18 KiB) Viewed 635 times

Give it a bit more help by adding 1961 and I got this

Screen Shot 2023-11-24 at 3.54.30 pm.jpg (119.75 KiB) Viewed 635 times

Clearly it hasn't learnt an Australian classic yet.
Gave it an easier task - Karmann Ghia. No year entered, but it mashed together something close to a late 60's Ghia.

Screen Shot 2023-11-24 at 4.01.52 pm.jpg (106.51 KiB) Viewed 635 times

Try things like Dash Karmann Ghia, Interior Karmann Ghia, and the results are just more exterior shots of mutated versions of a KG.

It's early days, frustrating 80% of the time, and a great time waster.

[Errol62]

AI + smoke machine.....

Sent from my SM-G781B using Tapatalk

[funkyscooter]

Mate we all rely on the real AI (Andrew's Intelligence). Your knowledge and willingness to share is one of the reasons this site is so important

[Harv]

AI + smoke machine.....

Cheers,
Harv

[In the Shed]

Interesting work Scott,
I like some of subtle features in the early iterations but then it goes off the rails to the point of nah!!

Back in the day a team of designers would probably work weeks on clay models coming up with design concepts that took you a few hours. Technology hey……

Regards
Stephen

[Blacky]

The last thing I need is another reason to faff around on a computer…….

[funkyscooter]

The last thing I need is another reason to faff around on a computer…….

Then my advice is don't start. I thought it would be fun for five minutes and have wasted a day on it. I always tried to draw cars as a kid, but I can't draw to save myself (see image below). It throws some really strange stuff out, but sometimes there is a subtle element in the line of a panel or a change to the bright work that you look at and go - that might be interesting. At least it can help solidify ideas before you start cutting things up.

I'll stop fanning out on this with one final example which is not going to win any design awards (maybe the Homer Simpson award for car design). The final result is all arse about for sure, but it's based on the dodgy 1 minute sketch I created, and the prompts I feed it. Not an example of modifying a Holden at all, but interesting in its potential to generate ideas from some very basic information
This is the image I drew - finger on track pad drawing to influence the AI how I wanted to view the car.

Screen Shot 2023-11-25 at 7.06.00 am.jpeg (76.58 KiB) Viewed 589 times

Prompt is a green 57 chev in snow

Screen Shot 2023-11-25 at 7.06.57 am.jpeg (88.62 KiB) Viewed 589 times

[funkyscooter]

Merry Christmas to all, thanks for another year of in depth knowledge. I'll be down at the beach!

Chistmas at the beach.png (1.48 MiB) Viewed 382 times