2D Fan Art AI MilaAI - Fan art

[rezaf]

Thanks for spelling all that stuff out aDDont, at least to me, generating art is a case of the old "easy to learn, hard to master" saying. Basic image generation from a prompt is very easy, the only difficulty is figuring out what words in the prompt can prod the AI in the right direction.
When trying something a little more advanced, most stuff I try flat out doesn't work, works terribly, produces horrible results ... all that stuff. Took me a while to figure out Regional Prompter, for example, and 3/4 of the howtos I read or watched about it didn't work as advertised or at all. Talk about motivational.
I figured out some stuff, so that's at least something.

What do you mean when you say "make/draw the base picture" - do you create a crude drawing to guide img2img in a paint program via copypasta or just create a sketch? I think someone used a sketch to guide the generation in one video I watched, but back then I had a hard time getting ANY useable results, so maybe I'm misremembering. Didn't get that to work, in any event.

That said, with Mila I found out today it makes a huge difference if you reduce the output resolution. On 1024/768 or even 800/600 it's impossible to get the "Mila look", or so it seems, even if it's a portrait or closeup where it would "fit", space wise. I usually get a much more anime-style look, with bigger eyes and all that stuff. I figure that's a PONY thing cause anime is it's forte.
I used your advice to use img2img instead and just used that same picture I mentioned as a base, and while the generated image has quite a few issues (I thought they had solved the hands thing at some point - apparently not), I think it's a solid start and recognizeable (I hope) as Mila.
I made another picture with Mila doing Yoga in the park, and I think the result is kinda sorta okay.
Am I on the right track?

 

[aDDont]

Thanks for spelling all that stuff out aDDont, at least to me, generating art is a case of the old "easy to learn, hard to master" saying. Basic image generation from a prompt is very easy, the only difficulty is figuring out what words in the prompt can prod the AI in the right direction.
When trying something a little more advanced, most stuff I try flat out doesn't work, works terribly, produces horrible results ... all that stuff. Took me a while to figure out Regional Prompter, for example, and 3/4 of the howtos I read or watched about it didn't work as advertised or at all. Talk about motivational.
I figured out some stuff, so that's at least something.

What do you mean when you say "make/draw the base picture" - do you create a crude drawing to guide img2img in a paint program via copypasta or just create a sketch? I think someone used a sketch to guide the generation in one video I watched, but back then I had a hard time getting ANY useable results, so maybe I'm misremembering. Didn't get that to work, in any event.

That said, with Mila I found out today it makes a huge difference if you reduce the output resolution. On 1024/768 or even 800/600 it's impossible to get the "Mila look", or so it seems, even if it's a portrait or closeup where it would "fit", space wise. I usually get a much more anime-style look, with bigger eyes and all that stuff. I figure that's a PONY thing cause anime is it's forte.
I used your advice to use img2img instead and just used that same picture I mentioned as a base, and while the generated image has quite a few issues (I thought they had solved the hands thing at some point - apparently not), I think it's a solid start and recognizeable (I hope) as Mila.
I made another picture with Mila doing Yoga in the park, and I think the result is kinda sorta okay.
Am I on the right track?

You're definitely on the right track.

Mila’s more anime-like appearance comes from using older-style prompts. In the latest releases, I’ve shifted to a slightly different style and adjusted the weight distribution of tags accordingly. The updated prompt is available in the pinned messages on Discord (accessible to everyone).

That said, even this style can be refined to match the usual look. You need to use a refiner with ponyfaetality and only masked inpainting on the face. In general, detailed post-processing is the longest and most tedious part, but it’s exactly what makes an image stand out from a standard AI-generated one.

I also want to emphasize that I use Photoshop a lot—I adjust composition and body proportions, for example. If Mila’s head is smaller than a 1:7 ratio compared to the rest of her body, AI will start turning her into a loli, which we don’t want (her ratio is around 1:7~7.5, Liz is 1:7.5~8). So, I either scale up the rest of the body or shrink the head before the next generation pass.

Regarding sketch preparation, yes, lately I’ve been using references or drawing a rough layout myself—symbolically marking objects, perspective, lights and shadows and composition—then gradually "pulling" the image in the direction I want. The process is similar to painting with blotches—it gives way more control over the final result and helps maintain consistency in costumes and environments.

 

[rezaf]

Thing is, if I use the prompt from the discord, at least in txt2img, I run back in the "not mila" problem...

 

[somguy]

Thanks for sharing the lora, I only just saw this now!

Here's some I generated, used ponyFaetality.

Spoiler: Pics

You don't have permission to view the spoiler content. Log in or register now.

 

[rezaf]

I don't know what exactly my problem is, but if it wasn't obvious before, somguy's excellent renders show the issue is something on my side. Maybe it's situational ... I did a couple of re-renders (unfortunately 1920x1080 renders take a good while to finish on my rig, should probably get a new gpu at some point) and this one is better, I think:

Edit: Can you share the prompt you used to create Lee (the blowjob guy from the park)? Might be a good candidate for some expansive renders.

 

[swiver]

You're definitely on the right track.

Mila’s more anime-like appearance comes from using older-style prompts. In the latest releases, I’ve shifted to a slightly different style and adjusted the weight distribution of tags accordingly. The updated prompt is available in the pinned messages on Discord (accessible to everyone).

That said, even this style can be refined to match the usual look. You need to use a refiner with ponyfaetality and only masked inpainting on the face. In general, detailed post-processing is the longest and most tedious part, but it’s exactly what makes an image stand out from a standard AI-generated one.

I also want to emphasize that I use Photoshop a lot—I adjust composition and body proportions, for example. If Mila’s head is smaller than a 1:7 ratio compared to the rest of her body, AI will start turning her into a loli, which we don’t want (her ratio is around 1:7~7.5, Liz is 1:7.5~8). So, I either scale up the rest of the body or shrink the head before the next generation pass.

Regarding sketch preparation, yes, lately I’ve been using references or drawing a rough layout myself—symbolically marking objects, perspective, lights and shadows and composition—then gradually "pulling" the image in the direction I want. The process is similar to painting with blotches—it gives way more control over the final result and helps maintain consistency in costumes and environments.

ever considered making video tutorials? would follow your channel!

 

[somguy]

I don't know what exactly my problem is, but if it wasn't obvious before, somguy's excellent renders show the issue is something on my side. Maybe it's situational ... I did a couple of re-renders (unfortunately 1920x1080 renders take a good while to finish on my rig, should probably get a new gpu at some point) and this one is better, I think:

Edit: Can you share the prompt you used to create Lee (the blowjob guy from the park)? Might be a good candidate for some expansive renders.

I was having similar issues to begin with, I couldn't generate a good image using the full sample prompt in the pinned comment, maybe I was missing a resource or something?

If it helps, here's my full generation data for the first pic.

Spoiler: Generation Data

You don't have permission to view the spoiler content. Log in or register now.

I uploaded the raw PNGs with full metadata but it looks like the site compresses every pic you upload to AV1, so you lose the metadata.

The prompts are based on some I found on discord (credit to User_P), I picked ponyFaetality for the checkpoint.

My usual process in general is render at 1024x576, upscaled with hires-fix to 1440p (2.5) or 1.875 for 1080. I use Forge instead of A1111 because it more efficient, A1111 would usually max out VRAM and take forever to do 1440p. I've had issues using larger resolutions in the past, like it puts 2 characters or you get 2 heads or other weird stuff lol, so I just render 1024 max and upscale now. Don't know if that's just a me problem?

Here's another with cyberrealisticPony for a more 'realistic' look. Nice but doesn't fit the style of the game (which I prefer tbh)

Spoiler: Pic

You don't have permission to view the spoiler content. Log in or register now.

Good luck!

 

[aDDont]

I was having similar issues to begin with, I couldn't generate a good image using the full sample prompt in the pinned comment, maybe I was missing a resource or something?

If the elongated limbs are bothering you, it’s most likely due to the aspect ratio. Pony doesn’t handle 16:9 very well, so I don’t recommend generating base images in that resolution. Instead, use 3:2, 4:3, or 1:1, and then either crop the excess or paint in the missing parts in Photoshop.

Another way to fix this is by selecting ResizeAndFill in img2img generation—it should help maintain better proportions.
a
If it’s the stylistic differences that are bothering you, again, it’s mainly due to the ponyfaetality refiner model. I also do a lot of only masked inpainting, especially on the face.

If you don’t want to spend too much time on manual fixes, there’s a "lazy" workaround—use adetailer with a separate prompt for the desired emotion and eye color. This should get you an image much closer to what you see in the game.

Also realistic Mila is hot too)

ever considered making video tutorials? would follow your channel!

With my rrrrussian accent my speech will sound like swearing
Tbh the hardest part in it is finding time, I spend the whole day on the last tutorial I made and I am not really a public kind of person) Maybe when I upload the songs on youtube I will also add some tutorial videos too, we'll see)

 

[swiver]

Also realistic Mila is hot too)

Really? The whole thing? Awesome, I'd love to replay the whole thing in this style! Good that you're going to wait till after you're done with the current look, though, because grrrrr.

I am not really a public kind of person

So you're more a bringing-you-orgasms-from-the-backstage/shadows type of guy, huh?

 

[aDDont]

Really? The whole thing? Awesome, I'd love to replay the whole thing in this style! Good that you're going to wait till after you're done with the current look, though, because grrrrr.

It was a comment about Mila in realistic style) I am playing with videos geneeration this week and there is (almost) no control over style yet. So I will make a couple examples in the realistic style, but I don't have plans of changing the one we have in the game right now.

 

[swiver]

It was a comment about Mila in realistic style) I am playing with videos geneeration this week and there is (almost) no control over style yet. So I will make a couple examples in the realistic style, but I don't have plans of changing the one we have in the game right now.

I know, perfectly understood you!

You meant after finishing the game you're going for a revamp in the realistic style, because there's no such thing as too much Mila

 

[rezaf]

Much thanks again to somguy, I think key was using the ponyFaetality checkpoint all along.
I'm still just playing around - I stuck to a small resolution today because that generates much faster on my rig.
Kept it pretty tame, too. But I do feel Mila is recognizeable on at least some of those.

 

[swiver]

Much thanks again to somguy, I think key was using the ponyFaetality checkpoint all along.
I'm still just playing around - I stuck to a small resolution today because that generates much faster on my rig.
Kept it pretty tame, too. But I do feel Mila is recognizeable on at least some of those.

 

[somguy]

Much thanks again to somguy, I think key was using the ponyFaetality checkpoint all along.
I'm still just playing around - I stuck to a small resolution today because that generates much faster on my rig.
Kept it pretty tame, too. But I do feel Mila is recognizeable on at least some of those.

Great work, these look amazing

 

[aDDont]

Couple examples of video generations. I don't plan to work with video for now - I am not satisfied with quality)

 

[DomSoul]

Couple examples of video generations. I don't plan to work with video for now - I am not satisfied with quality)

I have a different view on the videos being bad. I loved them!!! Epic!

 

[DomSoul]

Would love to see a cumshot in that video with her opening her mouth. The tech goes fast now. Soon you would get all them nice things in epic games