AI Image Generation - How to Thread

[anthony26]

I would like this thread to be about what are the latest AI Image gen tools people are using. I noticed they are all cracking down and even when not using NSFW words, It is trickier and tricker to get SW content.

Let's all discuss free or paid AI image gen tools. What works, what doesn't work. You can post your images as well as the prompts used and tool used to generate as well - so we can all become more proficient at AI Image Gen.

[Prof Sai]

https://apps.apple.com/us/app/draw-thin ... 6444050820

If you are on a PC, I can't help you, but there is probably something similar.

[Nonsensical Nano]

If you are on a PC (Windows/Linux), and your GPU is from Nivida and has 8GB or more RAM, https://github.com/LykosAI/StabilityMatrix .
StabilityMatrix is a tool that lets you download all kinds of AI tools for image generation and keeps them up to date. I used NMKD before, but using Stable Diffusion in a Browser GUI gives you much more control + you can get Lora/Lycoris/Checkpoints direct from CivitAI. There are two Loras/Lycoris especially for creating shrunken/giant things.

[Prof Sai]

iPhones and iPads can run DrawThings. That might be cheaper than a fancy PC graphics card. I'll tell you in a bit how well that works on my older iPad air. On M-series Macs, all ram is V-ram, so no need to worry about that. 16GB is still quite a bit better than 8GB, however.

Update: My A12 iPad with 3GB of ram can actually run Draw Things. But only at low resolution, 512x512, maybe a bit higher. And at least a minute per image. Not obviously useful.

[anthony26]

Prof Sai, I can't get DrawThings to give me good images. Can you give us a tutorial on what settings to use? I really like handheld shrunken woman images. I can't get it to produce anything good. I can't even get shrunken woman. I must not be using the settings properly.

[Prof Sai]

What got me stuck for a long time was trying to use the Generic model. I think the other models depend on it for some reason, but on it's own it is almost useless. You can download models from that menu, or look here for one you like:

https://civitai.com/

There are also LoRAs and Checkpoints, which I don't find very useful, and SDXL, which is good for landscapes, but not as good for people. I can't get SD to do shrunken women directly, but it is good for making elements to use in a collage. You'll need an art program to fix details and produce certain ideas that the AI just doesn't understand.

When experimenting, start with low resolution (512x512) and around 15 Steps. Those settings will render quickly, and help you learn what other settings do. There are still controls that I don't understand, and some like Face Restoration that unfortunately don't seem to do much of anything.

AI art is advancing very quickly, so all programs are rough with some clumsy interface elements. You just have to experiment, and do lots of renders to figure out how things work.

[foreverlurk]

I could write a lot on the subject but each tool has different strengths and weaknesses, and there is no single monolithic workflow that magically works for everything, it changes depending on which software you use and what exactly you are going for.

I use DALL-E through their API, with ChatGPT and through Bing Image Creator... it's confusing as hell, but those three methods are all different in subtle ways, Bing being the best IMHO for realism and getting past the filter, ChatGPT being the beast at generating prompts (which can then be used in Bing)

For SD, I use a mix of 1.5 and XL. I still use NMKD for simple stuff, but ComfyUI is much more powerful and comes with a steep learning curve, too. Sai is right about the default model, it's no use since it was trained on inferior source data compared to the specialized ones on Civitai. Loras, Lycoris and checkpoints all have different uses which would be too long to explain here. SD can actually do pretty good SW if you brute force your way through it (admittedly, having a 4090 helps when generating batches of 100s).

As for specific prompts, they change a LOT depending on the context. SD's language model (CLIP) is vastly inferior to ChatGPT's (used by DALL-E), so getting it to understand your prompt requires more work. For size stuff, the usual "low angle" or "high angle" perspective works well, but it doesn't do fisheye lenses or depth of field as good as DALL-E. I also found that "macro photography" triggered the right perspective sometimes, YMMV.

Here's a simple SD example with the popular Realistic Vision V5.1 model, made in NMKD 1.11

First pass
Resolution: 768x768

Prompt:
macro photo of very tiny, shrunken nude woman walking on kitchen countertop, filled with cookwares, lilliputian, short, shrinking, 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3

Negative Prompt:
(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers)1.4, (deformed, distorted, disfigured)1.3, poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, amputation

Steps: 20
Scale: 6.00
Seed: 735725126
Sampler: DPM++ 2M Karras

Second pass (img2img)
Resolution: 1472x1472
Scale: 5.00
Denoise : 0.4
Same seed, prompt, etc. except I added "smiling".



One day maybe I'll to write full tutorials, but it's moving so fast I wonder if its even worth it (ie. the info would be obsolete in weeks or months).

[Prof Sai]

The face is very good and hands are correct. But she has a bit of a plastic-doll sheen to her. Maybe their AI switches to CGI for things like this?

[foreverlurk]

The face is very good and hands are correct. But she has a bit of a plastic-doll sheen to her. Maybe their AI switches to CGI for things like this?

You are correct. Oddly enough, it's more apparent in Stable Diffusion because DALL-E 3 isn't that photorealistic to begin with, so those details get lost in the overall render. SD is vastly superior for realism, but from a certain size range, the model must infer data from plastic dolls and toys, thus the shiny look of the skin. What works to correct that is either inpainting, or img2img with a slightly different prompt and a lower denoise value (side effect of getting more randomness, though).

I've been working on another version, I think it's a bit better, or at least more natural-looking, let me know what you think :


You don't have that problem at bigger sizes, like around 3 feet tall:

[anthony26]

foreverlurk, do you have any recent ones where you were able to get handheld SWs? That is what I want to find a local tool to help me since bing doesn't seem to allow it anymore.

[Prof Sai]

This is an excellent image, very realistic, and something I can't achieve from DrawThings without substantial post-work.



A strange thing I have noticed. The vast majority of glamor models and nudes are shot with the camera at the level of her belly button, and looking up to her face? Why is that? If this was all about "male gaze" they would show women as men see them: The cyan horizon line would be at her eyes or the top of her head. Instead we always look up at the woman's face - wrong for SW art.

This shows up in AI art, as I had a great deal of difficulty getting images that didn't place the viewer below the subject. You can't really fix this in a 2d editor, as it makes the pose not match the background.

[foreverlurk]

foreverlurk, do you have any recent ones where you were able to get handheld SWs?

I'm not sure what you mean by recent ones, handhelds haven't really changed on DALL-E and they're much easier to generate than with a local Stable Diffusion instance. If you're talking about the filter aka. the dog, yeah it's more sensitive than before but if you're patient you'll get past it eventually.

Instead we always look up at the woman's face - wrong for SW art.

You're right, and it's annoying. I don't know if it's a bias of the training / source data? Or a side effect of altering the size/perspective? The models I use all have the same issue (Realistic Vision, in this case). As a workaround in the prompt I sometimes use "high angle photo of.." at the beginning, but you have to be lucky to get a good shot:

[Raso719]

Has anyone tried to use AI to make microscopic women? Either imposing women into views from microscopes or to try and create those views purely from AI?

[Prof Sai]

There we go!

I've noticed this long before AI started, so it is definitely an issue with training data.

[anthony26]

Agreed

[Pacific1982]

I’m here because I thought I would share my current results with AI. I’ve been using the free versions of DALL-E3 and ImageLabs Editor. That latter allows for NSFW and porn (and in fact brands each image with the word PornLabs.net). When trying to make shrunken women I get mostly deformed or misshaped creatures. Sometimes they look like dolls (Image 06 in particular). I thought the one riding the giant eagle was nice. #05 is supposed to be a normal-size college girl meeting a troll in the forest (á la Tolkien’s legendarium).

Should mention that I now use AI mainly for generating background images or normal-sized nude women, and then manipulate and/or incorporate those into SW creations as I see fit. I have also noticed that, too many times, the AI will omit one or more elements I ask for. And in several attempts, when I ask for full-figure view of a woman, I get only the head or upper body.

Thoughts?

[WhoIsOutOfOrder]

I’m here because I thought I would share my current results with AI. I’ve been using the free versions of DALL-E3 and ImageLabs Editor. That latter allows for NSFW and porn (and in fact brands each image with the word PornLabs.net). When trying to make shrunken women I get mostly deformed or misshaped creatures. Sometimes they look like dolls (Image 06 in particular). I thought the one riding the giant eagle was nice. #05 is supposed to be a normal-size college girl meeting a troll in the forest (á la Tolkien’s legendarium).

Should mention that I now use AI mainly for generating background images or normal-sized nude women, and then manipulate and/or incorporate those into SW creations as I see fit. I have also noticed that, too many times, the AI will omit one or more elements I ask for. And in several attempts, when I ask for full-figure view of a woman, I get only the head or upper body.

Thoughts?

I use stable diffusion, and you can use negative prompts there, for example "doll",figurine","toy","plastic" which gave me some amazing results.Negative prompts can sometimes even help with strange hands or body parts. "no doll","no toy" could help

[foreverlurk]

When trying to make shrunken women I get mostly deformed or misshaped creatures. Sometimes they look like dolls (Image 06 in particular).

Currently, all generative AIs will have the same problem with distant (or "shrunken") models. When they look like dolls but the rest of the scene is okay, we can use inpainting to render a better version. I have given it a try on my local SD install :



This requires a specialised inpainting model version, in this case I used Realistic Vision.

[Pacific1982]

Played around with one of your images (Foreverlurk?)

[foreverlurk]

Played around with one of your images (Foreverlurk?)

Nice work, Pacific1982! Love the contact shadows on her skin.

As you know, hands are the nemesis of AIs currently, so while I'd like to include them in 90% of my gens, sometimes it's easier not to.

[Pacific1982]

Foreverlurk, your inpainting improvement sounds intriguing, although perhaps a bit too much for me to chew on at one time. I took the trouble to isolate and enlarge the before and after of those two SW renders so that I can see them better. All I can say is, wow. This idea warrants further investigation on my part, time permitting.

[DocRick]

My valiant effort.......

[Prof Sai]

Aww, poor thing. I'll make you big again.

[foreverlurk]

Foreverlurk, your inpainting improvement sounds intriguing, although perhaps a bit too much for me to chew on at one time.

It's not too hard to use, but in my case I'm using a local Stable Diffusion install. I don't know if online SD providers offer that possibility?

I could have even better results than this, in terms of fine details and resolution, but it's always a tradeoff between accuracy (ie. how well it "fits" with the original pic, in terms of lighting, colours, etc) or better resolution and fine details. In this case, when the SW is small enough in the scene there's no need to go too crazy with the level of details.

Aww, poor thing. I'll make you big again.

Size restored!

[foreverlurk]

My valiant effort.......

Nice, now I'm thirsty lol!

You know I could generate those with a green-ish background or gradient to make it easier to cut for collages and stuff...