AI Artwork Tutorials, Tips & Tricks (Mega-Thread)

[JaxsonStark]

-__- Can't even get away from Ai "Art" in my porn.

if you want it that bad just learn to make morphs,
you'll get better results and its practiaclly the same thing since the ai is just taking from other pregnant pics and morphs and slotting them together.

why people think this is so great is beyond me.

I mean AI images are an endless avenue of new content. Also it has it's own section away from your preferred method of consumption. Morphing is a time consuming process if you actually want solid results, I do them myself and can say typing in a prompt and letting my PC run images in the background while I do other things is far easier than spending a few hours in Photoshop. While I don't understand your negativity towards this you are entitled to your own opinion, but your comments just seem so out of place and contribute very little to a thread which provides a lot of new and fresh content to the forum. ?

[User 70749]

Pretty sure they're just trolling, otoh some people are gonna be more sensitive to the inherent weirdness of AI art to the point that it's a dealbreaker. Believing that's the objective pov is not a opinion relevant to this thread.

[pregphil]

-__- Can't even get away from Ai "Art" in my porn.

if you want it that bad just learn to make morphs,
you'll get better results and its practiaclly the same thing since the ai is just taking from other pregnant pics and morphs and slotting them together.

why people think this is so great is beyond me.

Feel free to not post in this section.

[User 27138]

Those don't look too bad, but some people are more sensitive to the uncanny valley than others. 

The diffusion process relies in part on randomness so you're always going to be picking out the best of what gets generated. For everything you see shared there's a ton that get discarded for good reason. 

What image size are you making? The diffusion models are trained at 512x512, the taller the output size the more likely it is that the subject will become a distorted monstrosity. The Automatic1111 implementation has a fix for that which basically automates the process I follow for higher resolution images - creating a smaller image then running it through img2img. It's great but it's not selective and will just run every iteration through the process, which can waste time.

There's also ways to redo just parts of an image and you can touch things minor flaws up in an editor before sending it back through img2img. For that approach it really depends how much work you want to put in.

Using an implementation that lets you use negative prompts helps. It doesn't fix everything and if overused becomes counterproductive, but it can reduce the probability of anomalies.

I'll have to check out Automatic1111, I've seen that name elsewhere also, so there must be some useful techniques there. I'll try 512x512 my next run. I had height in the 700s.

How much does img2img really change things? Wouldn't it resulting image come out the same with the same tags and everything?

Here's one I did yesterday that looks alright but not exactly the way I want, especially with her shirt over her belly. Some people make really good stuff that's just crazy to me. I want to do that.

[User 70749]

For that one you can use the same seed at the same resolution and alter the prompts to change what Tifa is wearing. The seed is what determines the pseudorandom visual noise that is used to begin the diffusion process, so the same seed will give similar results for the same resolution, but you can't use the same seed on different resolutions.

img2img is too hard for me to explain well, if the settings aren't right then nothing changes OR the wrong things change OR there's too much change. There's no one size fits all, it depends if you're trying to use it to fix things or trying to upscale but with added detail and definition which non-commercial upscalers don't do.
 Without a NovelAI account I can't give much advice about their I2I interface if they have one.

Here's a link with the fork of Automatic1111 that I have been running on colab.

[User 70749]

Interface - Fast Auto1111 -  https://github.com/TheLastBen/fast-stable-diffusion 

Model - SD 1.5

Prompt:Beautiful curvy cute young teen country girl age 19 standing in the shower, very pregnant belly, large cleavage, thick thighs, naked and seductive, bellybutton,
Negative prompt: deformed, lumpy, bizarre, weird, amputee, extra limbs, weird hands, plastic, huge hands, deformed hands, anime, cartoon, no bellybutton, multiple bellybuttons, grainy
Steps: 20, Sampler: Euler, CFG scale: 10, Seed: 222569307, Size: 512x512, Model hash: 81761151, Batch size: 4, Batch pos: 0

Selected from a batch of 512x512
Edited to remove the typical fingerstubs that occur, quick and dirty work with the clone stamp tool. then run through img2img with crop and resize:
Prompt: Beautiful curvy cute young teen country girl age 19 standing in the bathroom, very pregnant belly, large cleavage, thick thighs, bellybutton,

Negative prompt: deformed, lumpy, bizarre, weird, amputee, extra limbs, weird hands, plastic, huge hands, deformed hands, anime, cartoon, no bellybutton, multiple bellybuttons, grainy
Steps: 20, Sampler: Euler, CFG scale: 16, Seed: 1732918755, Size: 960x1024, Model hash: 81761151, Denoising strength: 0.2, Mask blur: 4

I don't know how that lingerie-jeans hybrid outfit is supposed to work but I'm not complaining.

By using the same seed, settings and prompts anyone can produce the same output at each step.

This implementation also lets you merge different models. Same prompts/seed as last time:
From last time, Stable Diffusion 1.5:
   
Waifu Diffusion 1.3:
   
Merged Stable/Waifu Diffusion 16bit at 0.25 (0=100% SD, 1=100% WD)
   
use img2img for higher res.
   
Using/converting to 16bit models runs so much faster on the colab I'm using as well. Some hardware actually works faster with the 32bit models.

[User 70749]

just a couple of extras from messing with the prompts. Trying to undress her with prompts alone ruins the composition, which is to be expected. We're going to need to use infill to see what's underneath.

       

[User 70749]

I haven't had any luck working on that last one to get them undressed, I suspect image editing will be required.

Also I posted this in my thread but it belongs here too:

Experimenting again with making waifu art photorealistic:
   
It's still very much trial and error, this is the best result so far. Using weights that are a blend of SD1.5 and the f111 helped immensely, even though the latter is geared towards adult content it's based on photos so works better for this sort of thing than vanilla SD 1.5. I didn't think I'd use it much because I don't aim to replicate porn but it's been great for giving a photographic boost to both nsfw and sfw artistic content.  There's a sweet spot for cfg vs. denoising strength vs. interations that improves the likelihood of a good transition but it seems to vary betwen inputs, there's no magic numbers but I can at least say what worked this time.   The euler scheduler seems most reliable. 

Parameters: Steps: 24, Sampler: Euler, CFG scale: 8, Seed: 1448501382, Size: 704x1024, Model hash: c276019b (SD1.5 x f111, 0.4 ratio IIRC, float16), Batch size: 2, Batch pos: 1, Denoising strength: 0.7 

 The prompts are too cringe to share sorry but ofc there's (((photo))) etc. in the prompt and ((((drawn)))) etc. in the negative prompt. And even then it's not necessarily a photographic-style output.

I've had success doing this using the vanilla SD 1.5 model in the past and even just using DreamStudio's img2img, but couldn't make it work reliably.

Doing the inverse - transforming a photo into anine styke art - is relatively simple, just use img2img and the Waifu or NovelAI models:
   

Quite often Stable Diffusion won't reach the desired result in one go, but using a less than perfect output with img2img takes things closer to where we want them to go,

[FranticScribbler]

Anyone got any so art apps for those on mobile like devices?

[User 70749]

DawnAI looks okay at a glance:

   
   
   

The lexica.ai site has a useable mobile browser interface, but the way it upscales works better for artistic creations than for photorealistic ones.