Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Robsoie]

A very interesting use of those AI generating image tools, as an inspiration for a 3D artist artwork in Blender :
https://www.artstation.com/artwork/lR2mqa
Check his "Who created this artwork" article for more information.

[Robsoie]

that image guide feature on the mage website can lead into some interesting result, as you can tweak the influence of it on the prompt, you can even mask part of the image you want to keep on the final result, etc....
By example, using this excellent artwork from Yang Zheyy
https://www.artstation.com/artwork/9kmeO

Tweaking things a bit and you can get quickly something without effort with some simple prompt (and copy pasting the negative prompt example from a few posts ago) and some slider movement

Spoiler (click to show/hide)

I can imagine some various IP problem and lawsuits are going to happen in the future, especially when people will start to sell "their" art, i wonder how it will turn out as it will not be easy to find the real artwork dropped into the AI.

[Loud Whispers]

Spoiler (click to show/hide)

[freeformschooler]

Spoiler (click to show/hide)

I laughed

[Robsoie]

A rather cool and fun trick i learned from some people after checking the discord for https://www.mage.space/

By example on a square picture, i paint quickly this basic mix of color (the intent is me wanting to having nature, a huge rock, a waterfall coming from it and someone in front)

Spoiler: really very basic, no detail at all (click to show/hide)

I then use this as the image guide.
I setup the prompt as

A Naiad coming out of a waterfall

nothing complicated as you see.

and the negative prompt as the classic

lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck

no addition or word removal for now, we'll see how the generation will go.

Spoiler: first result (click to show/hide)

That's something better than my basic colored shape thingy, but that's not satisfying for me.

Let's change the prompt a bit, after googling around to find an artist name that may have painted a naiad , i found something nice by John William Waterhouse, so let's influence the AI generation with that, hoping the people that trained the AI fed it with this artist too

painting of a Naiad in front of a waterfall by John William Waterhouse

not more complicated than the previous one but hopefully the artist mention will do the trick.

a few Rerun of that as it generated several not really good to me images, until

Spoiler: the AI made something actually interesting for my eyes, warning it's probably slightly NSFW (click to show/hide)

Now no Rerun anymore, i click instead on "Re-image" , that function will take the actually generated image and will put it as the guide image (so it replace now my colored shapes thingy), i leave the strength of the image guidance to 0.8 so the AI have more freedom to work around this new guide image
(the closer to 1 that slider is will mean the more freedom the AI will have, the closer to 0 will mean the AI will be more constrained to the image guide).

I need to re-add the negative prompt as for some reason it's emptied after clicking on Re-image.

After a couple of re run of this new setup, i got another that triggered my "this has good potential" eye

Spoiler: slightly NSFW too (click to show/hide)

Instead of Re-run, i Re-Image this one (so it become the new image guide), this time i lower the image guide influence a bit (as i want to keep that overall character pose, and just hope there's enough AI freedom left so it can improve that character)
again, i re-add the negative prompt

After several Re-run (still the same prompt) i finally got a much better (to me) AI generated image

Spoiler: better character and background, oh and it's SFW this time (click to show/hide)

From there i could Re-Image and paint a mask over the waterfall and re run it to try to get the AI improve it, but that's enough for now.
Pretty cool trick.

[Loud Whispers]

Super cool, you can also use this to make very rare pepes

Spoiler (click to show/hide)

[King Zultan]

That's some fancy looking stuff there, are any of these art AI things free or do you have to pay for all of the better ones?


Also what's up with all of that stuff in the negative prompt box?

[Robsoie]

The one i use ( https://www.mage.space/ ) is totally free, but you need to make an account (that is also free) if you want to have all the (very useful) features, but you can give a try without account to see how it works.
Though with how good the thing is and its growing popularity/server load i wonder how much time it will survive before going "paid for"

While the prompt is there for you to try to get the AI making what you want, the negative prompt is there for you to input words to define what you don't want the AI to generate (it will slightly influence it, it's not really obeying you to the letter), it's super helpful to avoid some nightmarish results (unless you want that)

For fun, and because Eve Ventrue is an amazing artist and makes awesome artwork (check https://www.artstation.com/eventrue ) that worked for several high profile movie, i toyed with a simple prompt

Code: [Select]

art by Eve Ventrue

That surprisingly is enough (you can help with some negative prompt too) to generate some rather cool results without the need of an image guide.
After that, tweaking the prompt a bit, trying to influence the AI to go Imperial Guards , i got some interesting stuff (with some use of Re-Imaging of course as described in the trick before)

Spoiler (click to show/hide)

[Robsoie]

Re-using the trick i mentionned in a previous post with quickly and badly drawn ugly color shapes

Spoiler: yeah, me great artist :D (click to show/hide)

And using this "image" as a base guide, then after some prompt tweaking and using some artist name for the style influence, tons of Re-run (until i got a result that was having some potential) and some Re-imaging to re-use that potential and improve on it.

Spoiler: with so little effort involved, it's cool and sad in the same time ... (click to show/hide)

[Strongpoint]

I suspect that drawing women is the most common use of this thing)

[Robsoie]

Well, men have been drawing women since the dawn of time

[TamerVirus]

I suspect that drawing women is the most common use of this thing)

You aren’t wrong!

[Loud Whispers]

Those are pretty good ngl

[Robsoie]

toyed a bit with basic color shape as guide image
This time

Spoiler (click to show/hide)

as a prompt i used

woman in armor, 40k warhammer, sister of battle, (white haired bob cut), digital painting, correct hands, detailled, (cinematic lighting), dramatic, burning city in the background, artstation, 4K, masterpiece, art by artgerm and greg rutkowski and alphonse mucha

From what people have been saying the () supposedly are for insisting for the AI to use, but i wonder if it's just a placebo effect and it does not really work, at least on mage.space , but just in case it really do something let's add some.
the

art by artgerm and greg rutkowski and alphonse mucha

is because before doing that prompt i checked http://lexica.art/  and noticed some images that style was similar to what i was trying had these 3 artists in their prompt, so why not.

for the negative prompt

lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, out of frame, extra fingers, mutated hands, poorly drawn hands, deformed, blurry, bad anatomy, bad proportions, extra limbs, cloned face, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, face out frame, head out frame, body out frame, out of focus, out of frame

Anyways, i was re-running a lot of time without obtaining anything that was remotely satisfying to me even as a potential replacement for the image guide.
So i decided to edit my basic color shape image and mouse draw some small details in the hope it would give some more hints to the AI

Spoiler (click to show/hide)

And it worked, though it took a lot of Re-run to trigger my interest in the results, and delivered some image that i could use as very good new image guide to try to generate better stuff.
for those images , i had set the image guide "prompt strength" slider to 0.7 and the prompt "guidance scale" to 12.5

Spoiler: no result was holding some kind of future gun but at least it did not look ugly (click to show/hide)

Some observations : getting colored eyes is near impossible, even when added

red eyes

with or without () , the AI insisted mostly in doing regular colored eyes, it only very very rarely came to be red
Of course those damned AI hands, such a giant pain to get them not look like absolute horrors regardless of prompt.

something very annoying i also noticed is with the inpainting that allow you to paint/select/mask part of a picture guide so the AI would work only on the part you selected : the AI will STILL rework some part of the image despite your mask, check your character eyes by example after some inpainting the hands only, it's often infuriating

[lemon10]

I don't know about mage.space, but the ()/{} do matter depending on what site/webUI you are using.
The Novel AI website uses {} for emphasis with {dog} for instance having a emphasis of 1.05 and {{{dog}}} being 1.05^3=1.157
Stable diffusion uses () for emphasis with (dog) being 1.10 and (((dog))) being 1.1^3=1.33. [] is negative emphasis, with [dog] being 0.9

Since many people are using one of the two above most posts include some emphasis stuff.

Note that sometimes 1 emphasis isn't enough if what you are trying to get is particularly funky or conflicts with the picture you already have, and you want 2 or even 3 of them. Too many breaks the picture though, so watch out for that.

Below is an example of "cat hissing at dog" (done on Stable Diffusion using the NAI model) with different amounts of emphasis on dog. Note how the AI has a lot of difficulty with multiple focus objects and has them being the same thing.

Spoiler (click to show/hide)

Also here is an example of your prompt, but with different strengths of red eyes. As you can see once red eyes gets powerful enough funky things start to happen.

woman in armor,red eyes, 40k warhammer, sister of battle, (white haired bob cut), digital painting, correct hands, detailled, (cinematic lighting), dramatic, ((burning city in the background)), artstation, 4K, masterpiece, art by artgerm and greg rutkowski and alphonse mucha, holding sword in right hand

Spoiler (click to show/hide)

---
A minor note, apparently repeated negative tags don't really help and make the other negative tags less powerful, so

lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, out of frame, extra fingers, mutated hands, poorly drawn hands, deformed, blurry, bad anatomy, bad proportions, extra limbs, cloned face, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, face out frame, head out frame, body out frame, out of focus, out of frame

Should have worst quality, too many fingers, out of focus, and the second out of frame all removed at the minimum.

lowres, text, error, cropped, low quality, jpeg artifacts, ugly, duplicate, extra fingers, mutated hands, poorly drawn hands, deformed, blurry, bad anatomy, bad proportions, extra limbs, cloned face, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, face out frame, head out frame, body out frame, out of frame

You can probably remove other duplicates too, but I'm not 100% sure on the what the mage.space model considers important so I'm not sure if you really need all those out of frames+cropped, ect.