|
https://www.craiyon.com/ Basically I don't want to clog up good threads. So post what you got. Glory, to horror. If you've made a post in another thread, feel free to crosslink it. Soviet Hiluxen   next oh yes. oh poo poo yes.   Finally have some Group B rally cars: Ford Focus  Subaru Imprezza   And a Citroen C4 I guess?  Computers cannot do sponsor logos over the web I guess.
|
#
?
Sep 8, 2022 00:26
|
|
|
|
| # ? May 5, 2024 21:08 |
|
|
also stealing this shamelessly:mobby_6kl posted:It came out really well I think, as did all of KakerMix's images. A few minutes to clean up the extra headlights/wipers and it's a perfect Imperial 500.
|
|
#
?
Sep 8, 2022 00:31
|
|
|
Failure 1: entering hell backwards on fire in a honda civic painting kinkade  That said, excited for BMW's new American market only endeavor 
|
|
#
?
Sep 8, 2022 20:21
|
|
|
Still better than a new Bimmer looks
|
|
#
?
Sep 9, 2022 03:48
|
|
|
nanananananananana BEETLE!    Also I'm still trying to get a civic screeching into hell on fire painted by Kinkaid, but no dice  so I tried Bosch instead, but Dall-E seems to just choke on having Bosch in the prompt and keep creating essentially the same image 
Jonny Nox fucked around with this message at 08:11 on Sep 9, 2022 |
|
#
?
Sep 9, 2022 08:05
|
|
|
stiiiil on a beetle kick    kincaid status: nope  
Jonny Nox fucked around with this message at 08:37 on Sep 9, 2022 |
|
#
?
Sep 9, 2022 08:28
|
|
|
better.... 
|
|
#
?
Sep 9, 2022 18:05
|
|
|
It's no real suprise the Red Bull Toyota Prius racing series never took off edit: I'm trying to get a decent cab-forward ferarri, but it keeps making awesome images  
Jonny Nox fucked around with this message at 03:54 on Sep 10, 2022 |
|
#
?
Sep 10, 2022 01:56
|
|
|
Damnit people, diffusion is so slooooooow but so goooood  Group B Dodge Neon
|
|
#
?
Sep 10, 2022 07:00
|
|
|
Might as well get my Chrysler love out of my system:   1994 Neon 944  
|
|
#
?
Sep 10, 2022 07:58
|
|
|
Thanks for creating the thread, it does make sense to have some of these abominations separated  Stable Diffusion is quite a bit better than Craiyon, you can either run it on your own PC or use the free one here: https://huggingface.co/spaces/stabilityai/stable-diffusion although you have to wait in line so it's usually slower than my 1070. There's also a beta where you get a credit for a hundred images or so for free. There are also Dall-E and Midjourney but I'm not sure how they deal with casr. Widening the image to 768x512 seems to help keep the cars in frame. You can also steal prompts, like this one from KakerMix who had some incredible results: [year] [vehicle name], photo studio, ektachrome photograph volumetric lighting f8 aperture cinematic eastman 5384 film and change it as necessary the car description and location etc.
|
|
#
?
Sep 10, 2022 12:44
|
|
|
Trying some of Mobby's suggestions. Can't do 780x512 without more research though, I'm running a 1060 with 6GB VRAM Finally decided to just install the local version of Stable Diffusion. Next step is to figure out how to make more grimdark images. Adding warhammer and John Blanche doesn't have as dramatic effect on Stable Diffusion as it does on craiyon Either way, here's what happens when the Paris taxi lift goes too well  Also, Red Bull (at the time Redd Builll) absolutely clowned themselves at the 1954 Monaco Grand Prix   edit: Their sparse and poorly documented '70s efforts were equally poorly regarded   last one: 
Jonny Nox fucked around with this message at 09:40 on Sep 11, 2022 |
|
#
?
Sep 11, 2022 09:11
|
|
|
One this I will say about stable diffusion, it can do some Monet-rear end paintings  also kincade: 
|
|
#
?
Sep 13, 2022 15:24
|
|
|
Visions of the future:   edit: one I'm not going to press any further: 
Jonny Nox fucked around with this message at 00:55 on Sep 14, 2022 |
|
#
?
Sep 14, 2022 00:35
|
|
|
oh hello   
|
|
#
?
Sep 14, 2022 03:32
|
|
|
I made a shitload of ridiculous vehicle concept art in the style of syd mead concept art for a car with interior fire pit, illustrated by syd mead, high quality    concept art for a car with flamethrowers, illustrated by syd mead, high quality   concept art for a car that releases clouds of poisonous gas, illustrated by syd mead, high quality   concept art for bus helicopter, illustrated by syd mead, high quality    concept art for a car with a side car, illustrated by syd mead, high quality         that wasn't really what I was looking for, but I enjoyed the tiny cars so I tried to purposely create more concept art for a tiny car, illustrated by syd mead, high quality    
|
|
#
?
Sep 14, 2022 04:41
|
|
|
Stolen from reddit: 1970 BMW X7 Not gunna lie, I like it. Look at that glasshouse!
|
|
#
?
Sep 15, 2022 02:53
|
|
|
Dall-E has no idead what a SAAB looks like, but I got a pretty sick hotrod engine here. A SAAB hotrod with a chrome V10 engine. Color photo by Pete Biro. 
|
|
#
?
Sep 15, 2022 07:10
|
|
|
It's like all the Tri-Fives mashed up. BuckyDoneGun posted:Stolen from reddit: 1970 BMW X7 Whoa, that worked out well.
|
|
#
?
Sep 15, 2022 17:54
|
|
|
What promts are you using to get good crossbreed cars? Anyway, heres Ziggy Stardust as a hotrod. 
|
|
#
?
Sep 15, 2022 20:58
|
|
|
Honestly? It's been different for each one, and I'm not sure I'm really getting crossbreeds the way I want. VectorSigma is using img2img from SD (see second post in this thread or a link to install instructions) to get his images.
|
|
#
?
Sep 16, 2022 06:11
|
|
|
I'm using the Stable diffusion web version, since I can't run it locally on my laptop. Just putting the names of the cars together seems to work ok sometimes: "Ford V8 hotrod Fiat Topolino" turned out cute. 
|
|
#
?
Sep 16, 2022 08:53
|
|
|
 so close. yet so far. 1994 Ferrari Testarosa gullwing, studio photo, film grain Steps: 100, Sampler: Euler a, CFG scale: 10, Seed: 952177374, Size: 640x320
|
|
#
?
Sep 19, 2022 03:59
|
|
|
1985 Nissan Leaf
|
|
#
?
Sep 19, 2022 06:22
|
|
|
Kind of digging this Corvėtte DS
|
|
#
?
Sep 21, 2022 13:51
|
|
|
 a steampunk modified range rover from fury road that has a machine gun turret on the roof. post-apocalyptic land rover. shiny and chrome. war boys. mad max. photograph. film grain. volumetric lighting, F8, Kodachrome, wide angle. Steps: 50, Sampler: Euler a, CFG scale: 9, Seed: 3045575455, Size: 768x512, Denoising strength: 0.7
|
|
#
?
Sep 22, 2022 07:39
|
|
|
There's an new-ish UI with some cool features like negative prompts, built-in inpainitng, etc. Seems quite a bit faster (~20 seconds per image on a 1070) and can work with 8 or even 4 GB of VRAM. Easier to get working than the last one I tried too. https://github.com/AUTOMATIC1111/stable-diffusion-webui-feature-showcase Didn't have much time to mess with it yet so I just re-ran some of the old prompts. Miata pickup truck  WRC Daytona  80s sci-fi miata 
|
|
#
?
Sep 23, 2022 16:18
|
|
|
mobby_6kl posted:
Oh I absolutely adore this.
|
|
#
?
Sep 25, 2022 00:08
|
|
|
Jonny Nox posted:Oh I absolutely adore this.  But it doesn't seem to know him specifically so I'm getting more generic retro sci-fi type illustration which are also cool with me! The success rate isn't great but some cool stuff does fall out occasionally. I'm pretty mad the first one in particular wasn't wider (and outpainting fails completely)     If someone wants to teach the AI about their weirdo car (like, I dunno, the Traction Avant) it's possible to do it with just a few training images: https://textual-inversion.github.io/ I've tried it and failed so far, it needs quite a lot of GPU memory, though there's a way to optimize it to fit within 8GB supposedly.
|
|
#
?
Sep 25, 2022 11:14
|
|
|
Just gonna double post rudely because nobody else is posting like they should. I've been playing with the similar styles as before ("80s scifi illustration by shusei nagaoka, by Greg Rutkowski") but tried the RX-7. It seems to be a bit more confused about what it's supposed to look like but produces some cool results. I would drive all of these red cars (and one blue one)      
|
|
#
?
Oct 11, 2022 21:43
|
|
|
Is that last one an RX-7 F-body An... FD-body? 
|
|
#
?
Oct 11, 2022 21:51
|
|
|
Heck, that next to last one looks like a mashup of all three generations of RX-7. Would, all of them. Little bit of Cosmo 110S in that first one. Third one has a little Supra Mark III going on there. RX-7 with a quarter window (second pic) is novel.
|
|
#
?
Oct 12, 2022 21:15
|
|
|
Who knew concept artists would be the first to lose their jobs to AI.
|
|
#
?
Oct 12, 2022 22:03
|
|
|
RX7 Superbird
|
|
#
?
Oct 13, 2022 02:38
|
|
|
Yeah it's really weird. I did not specify a model name or year (it doesn't seem to work great, I'll post on that later) so it's a combination of the different generations and random 80s sports cars it seems. Here are the last two from that batch:  It can do an ND pretty well though  Can also do your shitbox too     
|
|
#
?
Oct 13, 2022 21:07
|
|
|
So, regarding the model model types. As mentioned before, I found it pretty difficult to get the specific one I wanted. It probably comes down to the dataset and how it was (or wasn't) labeled. Since it would mostly not be car nerds that use chassis codes, using an actual year seemed to work much better, but still not great. It can clearly distinguish the ND, and get the general vibes of the other models. It clearly can sometimes converge on an NA or NB (see below), but forcing it to do so is... challenging I ran some experiments, each row is 4 different results for the same prompt. Each row is a different model, specified by year or VIN code. Baseline & append model name  Append model year  Model name with underscore  Model year with underscore  Model year emphasized  Also another aspect are the styles. Some of them aren't really "neutral" and will tend to cause the model to converge on a particular style of car as well. I initially ran the experiment with the query I stole from someone here, "ektachrome photograph volumetric lighting f8 aperture cinematic eastman 5384 film, mazda miata". The results were vaguely miata-looking retro roadsters   It's possible to use weights on the year to counter it, but again see above for the limitations. Going to two or above serious wrecks poo poo lol.  As for the styles, this matrix is a bit messed up but clearly without "ektachrome" and "eastman 5384" it always draws an ND, but with them always an older NA-style roadster  Thanks for listening to my TED talk. As a bonus, here's a Civic 911, which is really what the Civic coupe should look like  
|
|
#
?
Oct 16, 2022 15:17
|
|
|
I'm really impressed with how accurately it draws the Honda and Mazda logos, I guess because they're single letters and distinctly letter shaped, rather than say the Chevy bowtie which isn't a letter, or Ferrari which is a stylized horse
|
|
#
?
Oct 16, 2022 20:07
|
|
|
mobby_6kl posted:So, regarding the model model types. As mentioned before, I found it pretty difficult to get the specific one I wanted. It probably comes down to the dataset and how it was (or wasn't) labeled. Since it would mostly not be car nerds that use chassis codes, using an actual year seemed to work much better, but still not great. So I've only done some light playing around with the public UI's, and being from a small place it's been really interesting trying to generate "local" content. I think the key thing is while we may not know exactly what datasets have gone into it, it's probably safe to assume that stock photography is a big part of it, possibly along with press photos. So, consider how those sets would be described and tagged - like you say, model years working but not chassis codes lines up with that. Back to the local content thing - "Auckland harbour" is a big area. We have two harbours with an infinite number of view points. However when I ask these AI's to generate Auckland harbour, they all look kinda the same - suspiciously like photos taken from the very popular and prominent viewpoints, which is also what you find in stock photos. Another one would be asking for a Māori person. In real life, not many Māori have the traditional moko face tattoo, they're regular people with brown skin. But AI's almost universally would generate someone with the moko! Again, because stock sets are full of Māori doing cultural displays so wearing full traditional garb. Imagine if every AI generated Japanese person was dressed as a samurai or a geisha - it's like that. I think it's an interesting observation on the limitations of the AI, it only knows what it knows and there's a strong element of paint by numbers.
|
|
#
?
Oct 16, 2022 22:30
|
|
|
Like "AI" for anything, it's down to it's training and dataset. There's massive concern in the popular art world (Think, ArtStation, DeviantArt, and other less family friendly image posting sites) regarding AI art. There's one AI in particular that uses on of the "boorus" (image posting sites, typically for NSFW stuff, often anime-style art) as it's source, as those images tend to be well tagged, and it's frighteningly good. You can still tell in a lot of cases (similar to wheels and badging on cars, AI has trouble with hands, feet, and some other stuff,) but with culls and refinements, some of it is very good. Which is bad, given how it's produced (stealing, basically, since this is actual art and creative work rather than stock photos and such.)
|
|
#
?
Oct 17, 2022 18:38
|
|
|
|
| # ? May 5, 2024 21:08 |
|
|
The dataset definitely has a huge impact on the results, both in terms of scope and tagging. Some concerns about diversity and bias got brought up almost immediately (for DALL-E?), like "doctor" was always a white dude, because that's what most doctors in the sample looked like. Actually tested this a while ago in SD (posted in the GBS thread), "woman in red dress" is surprisingly diverse, of different ethnicities and body shapes, and not just skinny white supermodels (though all conventionally attractive). Anyway, I believe I found the actual dataset that was used for training, and you can search it by keywords: https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn5.laion.ai&index=laion5B&useMclip=false&query=mazda+miata  Sometimes it's labeled MX-5, date is more frequent than the VIN codes, some garbage like stickers and shirts. Darchangel posted:Like "AI" for anything, it's down to it's training and dataset.  It's not great for cars, but can still do some cool artsy stuff    Anyway, back with vanilla SD:    It can absolutely nail the ND, and for whatever reason, that's the model it settled on 100% of the time for this style, even though I never specified it. Unfortunately though, the actual impressionist style doesn't seem to properly apply to the car, no matter how much I tried emphasizing it in the prompt. This was the case for all images in this style. I have another experiment running (100 images) so I'll go make dinner and then make a separate post about that (if it's interesting, hopefully).
|
|
#
?
Oct 17, 2022 22:26
|
|