Artwork Messing around with AI, A Art thread.

Yinko

Well-known member

Architects prove that their Avant-Garde style is now completely replaceable via AI. Traditional architecture requires that things make sense, that wall A connects to wall B and that the roof panels are orientated to provide good drainage. Post-Modern architecture... doesn't. So, that entire style is now trash. I mean, it's trash to the eye of the public as well as to the critics because it requires zero effort.

As an aside, I've been noticing on sites like Pinterest and ArtStation that pictures of attractive women that my brain suspects to be AI generated I have a substantial bias against. I wouldn't call it uncanny valley, since it doesn't look wrong. I don't know what it is exactly.
 

TheRejectionist

TheRejectionist
throwawayagainforsake_Candlelight_Dungeon_city_with_ghouls_seen_73fc0a7a-df7f-4176-9e50-57b0f40fbd48.png


throwawayagainforsake_Candlelight_Dungeon_city_with_ghouls_seen_0d46c4bd-33b1-4e6a-86c7-82f2867cc170.png


throwawayagainforsake_Candlelight_Dungeon_city_with_ghouls_seen_086aca38-01b6-4b35-952c-85f855d502c3.png

throwawayagainforsake_Candlelight_Dungeon_city_seen_from_above__e40ee306-f543-4b4b-a64b-c84b15df2401.png


throwawayagainforsake_Candlelight_Dungeon_city_seen_from_above__b727e5d0-0354-49c7-9dae-d9aafa7614b5.png


throwawayagainforsake_Mindflayer_beholder_crossbreed_Photograph_ac1d33b3-8605-49f3-bac9-14f8682e4c05.png

throwawayagainforsake_Butchers_black_glass_sword_Photography_hy_c5d1fc43-089d-4724-b2ce-ef2a04df4f1f.png

throwawayagainforsake_Obsidian_metal_sword_Photography_hyper_re_83c1f4ee-ffaa-4903-b19e-443924888d36.png

throwawayagainforsake_Obsidian_cutlass_Photography_hyper_realis_9d1e9329-7cac-4582-abc2-1f7005c787ec.png
 

Culsu

Agent of the Central Plasma
Founder
So, MJ v5 was recently released, and it's a mixed bag. The system's grasp of hands, and of holding things has vastly imrpoved compared to v3 or v4, and the level of detail it generates even on the base images is nothing but stunning. However, but by and large its still dumb a sa rock. It can't count, it can't really parse the meaning of language (what it does seem to do is latch onto the very first words of the prompt, ignoring or vastly downgrading the importance of everything that follows). Visual compositions that would be obvious/understandable to a pre-schooler are a mystery to it. As long as the devs don't work on the machine's understanding of language everything else is just window-dressing: yes, the system will create some stunning images; no, they won't really be what you've asked it to do.

Anyway, here's some examples of my rampage through v5.

Probably the most Battlemech-like mech it's ever manage to do for me:
0_0.png


Command center concept art:
0_0.png


A Shadowrun shaman (that's at least what I asked for).
0_1.png


Shadowrun troll merc:
0_0.png


Random Shadowrun corpo exec. She was supposed to be an elf, but MJ shows a major resistance to depict elves in anything other than the most bog-standard fantasy images, and even then it's very hit and miss. Used to be better.
0_2.png


Shadowrun cyber-samurai. Again, very hard to get the algorythm to mesh traditional Japanese samurai with the idea of cyberpunk. I think older versions had less trouble.
0_0.png


Again, the elven problem. Still, hands and objects are way better than in earlier versions.
0_2.png


A star sector map. v5 still is limited in its upscaling as its in alpha.
0_2.png


Could very well be a BTech tank.
0_1.png


Starship art.
0_2.png
 

Culsu

Agent of the Central Plasma
Founder
In case you're interested, MidJourney just unrolled a new feature that lets you upload an image and then has the AI create a series of prompts of that image. It's interesting insofar as you can clearly see that all your super long prompts are basically pointless; the AI only takes a few 'pointers' and runs with it. It's super obvious because none of the prompts the AI generates from provided images produce anything even close to the original image!
 

Culsu

Agent of the Central Plasma
Founder
Truth be told, as long as MJ doesn't use the same language algorythms like, say ChatGPT it'll always be "just" a somewhat random image generator.

On a different not, after ages of trying this is about as close as I've ever gotten with trying to have MJ come up with a spheroid BTech dropship in the style of the sourcebooks.

grid_0.png
 

Culsu

Agent of the Central Plasma
Founder
After having tried to get decent 'mechs out of MJ for months now I found out that the way to do it is to activate it's inbuilt anime style (Niji). Well, thanks for letting me waste my time, MJ!

Niji succeeds were vanilla MJ fails: missile tubes and guns, and distinct walker shapes. Only the visual fidelity varies wildly for no apparent reason whatsoever.

0_1.png



FWCarto_Mad_Cat_class_battlemech_clean_black_and_white_lineart__fd1af116-4589-4dca-a2c6-691cbca39ee1.png


FWCarto_Mad_Cat_class_battlemech_clean_black_and_white_lineart__a37606f5-c138-4a9b-b349-01df7f3308e2.png
 

Bear Ribs

Well-known member
I threw this brief tutorial together for Charclone in a PM, but I realized it might help people here too.



First thing to realize is that AI art is a bit of a gacha. We only publish our better images (usually) and there's several dozen failures for every one that's good. I usually set my system to batch-producing fifty to a hundred images overnight, and maybe one or two will be worth more than recoiling followed by the delete button in the morning. Those one or two will need a few rounds of inpainting and img2img refinement before I want to publish them. For every decent picture somebody posts there were five or six with eleven fingers, an arm growing out of a guy's belly button, or a girl whose mouth is on her throat instead of her face. For the masterpiece quality ones there were probably dozens, maybe hundreds of bad results and slow tweaks using inpaint to get them looking so good.

For getting you started, adding () around a tag increases its weight to the AI, its importance in the picture. So bunny ears is of basic weight while (bunny ears) means the AI knows that's of more important and (((bunny ears))) is given high priority. It's also a good idea to use synonyms, the AI thinks in keywords and models do not always use the same keyword for everything so throwing synonyms at the problem makes it more likely to work.

Note also that every tag affects the entire image, f'rex if you add Office background you'll get more bunnies wearing business suits even if that's not in your prompt at all. The AI is accustomed to seeing people in suits inside offices so it will make that mental connection for itself.

Bad prompt:
A bunnygirl wearing a military uniform and armor.

Better:
sharp focus,(8k), (4k), (Masterpiece), (Best Quality, detailed,) (rabbit ears), 1girl, military uniform, (breastplate), knight, armor,

Negative Prompt
(low quality:1.4), (worst quality:1.4),(monochrome:0.8),(extra digits), (cat ears), animal ears, (writing),(signature),(watermark),(words),

Using that Prompt I did a test run and generated a 12 image set. This was using the Anything V3 checkpoint for my base and not including any LORAs, to keep things simple.

R2wAQ12.jpg


In hindsight I should have added (Rifle), gun, gunbelt to it since you want those elements there. My mistake but I don't have time to do another run right now. This prompt was agnostic as to background, hair color, and facial expression. Assuming I had a more clear idea of what I was doing, I would also add things like blond hair, brown hair, or (blue hair) and smile, happy, angry, etc. to change the facial expression. This checkpoint seems to have a bias for putting blond or pale hair on bunnygirls so it would need weightier hair color tags to compensate if you want a redheaded warrior bunny.

Of those images, I decided to img2img this one:
DdB3tsY.png


It seems to have a slight sci-fi bent to the armor, and while the arms and hands aren't great they aren't mangled easier. Also, her head is lower in the frame making the ears more visible.

I added the (rifle), gun tags to it and it coughed up these img2img results, you can see there's some real improvement and a few more iterations on one of these could yield a halfway decent book cover, with some work refining specific areas and playing with the tags.
JC2acYg.png


Notice that even though I didn't add Gunbelt, the girls have them anyway. This is an example of what I meant by all tags affecting everything, because a large percentage of images with guns also have people wearing belts with ammo pouches or clips on them, it gave the Warrior Bunnies such belts just because there was a gun tag and it thinks those go together. Similarly if you add Motorcycle it's more likely to give them helmets, or if you put them in swimsuits the background is going to change to a pool or beach.

For these specific images, the rifle is quite mangled in one, and one has her fingers blending into the woodwork of the gun which would be really hard to fix. So assuming I kept refining them, I'd go with one of these.
mTqhbnr.png

SxPc1FH.png


Both have largely decent bodies and faces but would need so inpainting to fix the hands up and hopefully improve the guns. I like the first one better myself


Hope this helped, if you didn't understand any of the terminology or need more help let me know.
 

Users who are viewing this thread

Top