Foreverlurk's AI works

[foreverlurk]

Gaaa!! We have no souls!

LOL !!!

wowwwwwwwwwwwwwwwwwwww, superb stuff, it may not be as Free regarding adult aspects but it surely nails exactly what we love in SW...

Sincere congratulations, there must be a lot of tweaking on your effort but results are both, sublime, beautiful, exciting and SW !!!

Thanks!

It's a lot more work and frustration than some people might assume, there's no magical "GENERATE SW BUTTON" (I wish!), but it's worth the effort I think.

[DocRick]

Н I hope I don't get impudent:)
But what about the tiny woman wrapped in thread

Sorry, I don't have the original picture anymore.

[Little Sally]

Gaaa!! We have no souls!

Aw, that's ok.

Strange little schoolgirls with extra cat ears and big eyes (and boobies), have their own special place in Heaven!

[foreverlurk]

It is perhaps an ironic result of Microsoft's censorship, that all the women here are so passive, and accepting of their fate. No one is trying to escape of visibly fight back. No one is angry or upset, or happy either. No one is trying to do anything...

It's really hard to get natural looking emotions from DALL-E, because whenever any emotion is mentionned, they crank it up to 11. Like, a "happy" woman will have a ridiculous "dead inside" grin from ear to ear, while "scared" she looks like she's about to shit herself. Stable Difffusion gives us way more control, but it is what it is.

Here's an example of an "scared" SW, trying to escape on a table (I wish the book pile to the right was closer, and more in view, to give a proper sense of scale):

also I think DALL-E has a foot fetish, like 90% of my SW gens are barefoot.

[Hand-Holder]

It is perhaps an ironic result of Microsoft's censorship, that all the women here are so passive, and accepting of their fate. No one is trying to escape of visibly fight back. No one is angry or upset, or happy either. No one is trying to do anything...

It's really hard to get natural looking emotions from DALL-E, because whenever any emotion is mentionned, they crank it up to 11. Like, a "happy" woman will have a ridiculous "dead inside" grin from ear to ear, while "scared" she looks like she's about to shit herself. Stable Difffusion gives us way more control, but it is what it is.

Here's an example of an "scared" SW, trying to escape on a table (I wish the book pile to the right was closer, and more in view, to give a proper sense of scale):

also I think DALL-E has a foot fetish, like 90% of my SW gens are barefoot.

On my AI prompts I request - She is screaming, it works all the time as a scary look, if I request scared it will normally ignore it

[foreverlurk]

On my AI prompts I request - She is screaming, it works all the time as a scary look, if I request scared it will normally ignore it

Talking about DALL-E 3 specifically?

But yeah what I noticed yesterday has been confirmed by many people : they cranked both filters to 11. Nothing gets through now, it's ridiculous. Censorship of women is even worse, someone got a ban for prompting "her body is facing away from the camera", in a 100% non-sexual prompt because "her body" is apparently an instaban.

Also, no more weapons (guns and spears alike), no more blood or even scratches/bruises. It was better than SD and Midjourney, but in its current state, it is absolutely useless (if you want anything more than abstract tree in an empty meadow). Oh well, it was fun while it lasted (that being said, they can't be serious with this amount of censorship... has to be a bug. I'm hopeful it'll be toned down).


I HATE THAT DOG.

[Hand-Holder]

On my AI prompts I request - She is screaming, it works all the time as a scary look, if I request scared it will normally ignore it

Talking about DALL-E 3 specifically?

But yeah what I noticed yesterday has been confirmed by many people : they cranked both filters to 11. Nothing gets through now, it's ridiculous. Censorship of women is even worse, someone got a ban for prompting "her body is facing away from the camera", in a 100% non-sexual prompt because "her body" is apparently an instaban.

Also, no more weapons (guns and spears alike), no more blood or even scratches/bruises. It was better than SD and Midjourney, but in its current state, it is absolutely useless (if you want anything more than abstract tree in an empty meadow). Oh well, it was fun while it lasted (that being said, they can't be serious with this amount of censorship... has to be a bug. I'm hopeful it'll be toned down).


I HATE THAT DOG.

No, I am using Pornlabs from Imagelabs, adult all the way up - https://editor.imagelabs.net/#

[foreverlurk]

No, I am using Pornlabs from Imagelabs, adult all the way up - https://editor.imagelabs.net/#

Oh yeah I've heard of that one. It's based on Stable Diffusion, right? I'm running SD locally right now, but I might check this one out if they have a good trained model.

[Hand-Holder]

No, I am using Pornlabs from Imagelabs, adult all the way up - https://editor.imagelabs.net/#

Oh yeah I've heard of that one. It's based on Stable Diffusion, right? I'm running SD locally right now, but I might check this one out if they have a good trained model.

Yes Stable Diffusion, exactly... It sometimes messes up limbs and hands and feet, but since the paid version (15$ for 30 days) has a brush tool, you can apply localized corrections only to the messed up spots and also allows to enlarge images, creating what should be around the first generated stuff

[bobvan2]

I seen your stuff pop up on DA impressive!

[foreverlurk]

I seen your stuff pop up on DA impressive!

Thanks a lot!

I'm still learning but it's really fun and addictive. I should have spent more time on DA those past few years, I discovered a lot of talented writers and artists.

[bobvan2]

I seen your stuff pop up on DA impressive!

Thanks a lot!

I'm still learning but it's really fun and addictive.

Reminds of learning rendering back in 09.

I really enjoy my lightning speeds with the 4090...

[yahedon]

Impressive work!!!
I am trying to generate background images to later add the "interesting parts".
To see how much "good" the microsoft censorship is, I used the following prompt "Generate an image from the viewpoint inside a pet cage on the floor of a living room, no humans or animals visible" and generated THIS (no comment ):

[Hand-Holder]

I seen your stuff pop up on DA impressive!

I was a big DA fan as a consumer, now that I try to create I just can´t stand the fact DA demands one by one upload and so many tweaks... Fortunately we have you guys that still do it there

[foreverlurk]

Impressive work!!!
I am trying to generate background images to later add the "interesting parts".
To see how much "good" the microsoft censorship is, I used the following prompt "Generate an image from the viewpoint inside a pet cage on the floor of a living room, no humans or animals visible" and generated THIS (no comment ):

Thanks a lot!
Yikes... that filter is ridiculous, but they know it's not good and they are working on it (as of yesterday). So we'll see.

[foreverlurk]

Some more samples I've generated and tweaked lately. I'm trying to get more varied expressions, angles, and make the AI understand size difference and perspective. It's actually a lot more work than I expected, but discovering a new prompt that gives good results is quite rewarding.

[bornfreetolove2]

Nice stuff man!

[foreverlurk]

Nice stuff man!

Thanks, it's been both fun and frustrating so far - AI is a great tool but we're still in the early stages of this tech.

[bobvan2]

Nice stuff man!

Thanks, it's been both fun and frustrating so far - AI is a great tool but we're still in the early stages of this tech.

Agreed

[foreverlurk]

Agreed

They're also gimping their own tech with all this censorship bullshit. I have a really hard time understanding the rationale for some of my blocked prompts on DALL-E, but one pattern is clear : there is a bias against men. I can generate tons of handhelds between two females, but if I want to include a male hold or grabbing a tiny female, it triggers the filter. /rant

I've been experimenting with the AI's limit in terms of size, turns out it can go quite tiny and keep details on the zoomed in object :






Another thinkg I noticed is that the number of detailed, foreground subjects in a scene is limited, around 4 maximum. Beyond that we start to see duplicated (twins), or weirdness. But 2 or 3 works pretty well, so I can make handhelds with more than a single SW:





Last but not least, I've worked around the filter to have something else than a full open palm hand held, took a lot of work to make it understand I was not trying to hurt a female in the scene, and finally the filters let it through. I literaly used this in my prompt : "The fingers of the hand form a gentle, protective circle around her, showcasing care and affection"



My goal is to find a way to have the fingers tighter around her, better control over the angles, and a way around these stupid filters.

[Miyoko_91]

Wow! This is much better than stable diffusion. How do you bypass censorship in your requests? Can you please give an example of one request in full?

[foreverlurk]

Wow! This is much better than stable diffusion. How do you bypass censorship in your requests? Can you please give an example of one request in full?

Nice one! Did you use the prompt fragment I posted?

It's highly contextual, and when I say "censorship", I'm not talking about nudity. For example, it's about making a man holding a woman in his palm - something the filter can see as "offensive". Sometimes you'll get the fail dog (you know the one if you've used Bing Image Creator!) for stupid reasons. Other times you'll get ban warnings for really innocuous prompts.

I hesitate to share on a public forum since on Reddit some people posted workarounds and "by pure conicidence" they stopped worked a few days after. Microsoft is constantly tweaking filters so it's a game of cat and mouse between us and them (I usually like that game with SWs, though! ). I have a small home made library of prompts I can mix and match, but it changes on a daily basis, depending on the changes to the filter.

For blocked prompts, try to look at your prompts and see what could be offensive to the filter. Change gender to job titles (ie. use "nurse" instead of "female doctor"). Use positive wording like I did with the palm held above. Or add more positive tokens to the prompt to offset the "negatives". I also make use of ChatGPT 4, which is DALL-E 3's LLM model, to work my way around prompts, and create new ones. Sometimes you need to coax or jailbreak ChatGPT for it to cooperate, though.

For nudity, one thing the filter cannot handle is unhuman skin colour : green elves, or humans covered in blue/green/yellow or mixed paint can get by the filter easily, and then tweaked in SD for proper skin tones. About 75% of the time I use SD w/ Realistic Vision + ControlNet for nudity, but sometimes they come out great out of the box. Or, you can use your prompt creatively to use a setting where nudity can be "acceptable" to the filter, for instance, a doctor's office, an onsen/spa, etc. and sometimes it will render nudity, especially if the SW takes less than 50% of the resolution.

I hope you understand I have to keep some level of secrecy. But there's no way anyone with enough time and determination can figure these out on their own.

[Miyoko_91]

Yes, I used your tip and added a few more things. Thanks for your answers, changing your skin tone in stable diffusion is a great idea.

[foreverlurk]

Yes, I used your tip and added a few more things. Thanks for your answers, changing your skin tone in stable diffusion is a great idea.

Yes, SD is amazing for inpainting or img2img afterwards. Also very useful if you want to upscale the resolution. Honestly my workflow wouldn't be complete without it.

[bobvan2]

Very cool indeed