Requests for AI generation

[yano2mch]

This is intended as a general purpose thread, request something to be made and i'll run it through using my hardware. (Most of the time the video card isn't utilized enough...)

First preferences is to use smaller SD1.5 models, these are fairly fast and do a really decent job, just not at the highest resolution. These I've had good speed at good looking results.

Larger models take considerably longer but beyond less likely to have fucked up hands and feet and weird oddities. Regardless, the SD1.5 i can run like 50 images, in the time i get 1-3 images from the XL models.

So ground rules:

* I am happy to try and generate book covers for ones that don't have it. We're talking Literotica, your own writings, books that for some reason we have text-only format.

* SFW content i am happy to run through, there's a lot of interesting models

* For adult content keep ages 16+. I would rather just not have have stuff combed over.

* Don't expect super realistic results. Some of the AI art output i couldn't tell you they weren't AI art, but mine are probably going to be stylized or the like.

* For multiple characters interacting... it's hit or miss. No clue, i'm new to this.

* Don't ask for real people/characters. Example would be Luke Skywalker (Mark Hamil) doing anything. I suppose an exception to this would be if we made it hand-drawn or obviously stylized.

* Refrain from Disney and Nintendo properties, they are too lawsuit happy, easier to avoid.

* I'm fine with generating any species/races, assuming i can get a result. Though if we go furry, i'd prefer if it's for character profiles and not feel like i'm browsing FA.

* Not interested in extreme kinks, so no gore, asshole shots, or anything of those natures.


So, provide book/story references, name of the book and which chapter, etc. If you provide all the stuff even better than me fetching it.

Specifying the scene with as much reasonable (yet terse) detail as you reasonably can; Prompts need to stay under 300 words, or it won't do a good job.

Using Img2Img with a description, i can also heavily alter an image or have it reference it to redraw in a style or changed method. You can take a picture of yourself in a pose.

Most likely I'll do multiple generations at resolutions 768x768 and 1024x1024, and pick what i think is the best one(s), and in most cases probably refine the prompt. I'll probably dump most of the images in a mega folder otherwise to peruse.

--- new ones

 

[Bossyboots]

Arabian Harem. Two blondes naked

 

[Bossyboots]

A few more ideas...

 

[Bossyboots]

More in Bondage...

 

[yano2mch]

A recent model came out, qwen image generation. Fast, decent, not quite sure what the architecture is, but it's very promising for decent models at lower memory requirements. Some models wanted to take 30 minutes to generate a image, but this one about 20 seconds for 1024x1024, my cpu didn't barf so it somehow fits in 8Gb or it's doing a very good job of splitting the job.

Either way. I got the 8bit version which is 20Gb. Quite surprised... Maybe i'll find a good book on how LLMs and SD works, aside from the high-level overview. (Not that it will change anything)



edit: I'm not sure, i'm guessing my system barfed and these are likely the result of Dreamweaver or gonuts realism.. Figured like LLM's it would load what it could in the GPU, while now it seems more likely the whole thing needs to fit? I'm not sure.

 

[yano2mch]

Got a few models for pulp fiction. This is the hand-drawn style you see of a lot of older books of old and classics (Police academy comes to mind). Might consider doing them for making covers for substantial literotica or other books, rather than having a modern look.

Here's lady with parasol and lamia and bottle of whisky... Just for some tests. Though it looks like you'll get a lot more reliable pre 50's output than anything more modern.

 

[Haremfan]

Can you photo realistic these 2 images?

 

[yano2mch]

Can you photo realistic these 2 images?

View attachment 1850376 View attachment 1850378

Apologies, wasn't notified, guess when you make a thread you aren't automatically watching it.

First one i had marginal success, the second (though looking at them side-by-side they aren't that different), not so much. Would require a lot of inpainting and generation to probably get it halfway decent. Curiously i tried 50 steps at a lower (20-25% noise) and it kept most of the details.

 

[yano2mch]

A combination SD Pulp model and hand drawn lora has resulted in some very nice images that might be useful, be it for covers or illustrations. They have a nice balance of vintage while also semi-realistic.

 

[Cooffrree]

Have you tried to use pony or illustration(?) instead of SD15?

 

[yano2mch]

Have you tried to use pony or illustration(?) instead of SD15?

It's more a limitation of VRam than anything else. I've tried Pony/SDXL before, and had good results. But it starts demanding 5-30 minutes for 1 image, though maybe if I'm not using any LORA it may fit, but you really need 500Mb+ for the image being worked on.

That's a main reason I'm shy to SDXL, Pony and Illustration, but i get the feeling i'd have the same VRam problems.

edit: Ahh right, and also fighting with and trying to get image generations to work right, using clipskip and other features that aren't explained very well. The preview shows something beautiful but then i get this useless blob of nothing that you can vaguely see the shape of what it was making.

edit2: Yep... this is why i was giving up on those models. No clue. I don't want to spend 20 minute generations and take 10 hours trying tweaks when i don't know why it's coming out as garbage.



edit3: Not sure, but forcibly removing VAE's seem to have resolved most of the issues.

 

[yano2mch]

Well since 8Gb video card isn't going to cut it, ordered a 16Gb card that's a good balance of cost vs performance. Sooooo.... Maybe i can try the Qwen Image later. But for now be patient, and maybe i can offer the SDXL and other models when they actually fit in my card. (At present doing some generation with 6Gb models and no VAE and they sometimes work and sometimes stall... fun...)

@Cooffrree With a larger card Pony may actually be an option. I've only had a 8Gb card for... 6 months? And it was apparent 8Gb wasn't quite enough for these tasks, but i didn't think immediately buying another card made sense at the time. And now with Black Friday, well... We'll see

 

[alebeard]

I use a 16 gig card on my desktop but only an 8 gig card on my laptop and I run pony XL just fine using stable Diffusion Forge Neo

 

[yano2mch]

I use a 16 gig card on my desktop but only an 8 gig card on my laptop and I run pony XL just fine using stable Diffusion Forge Neo

Mhmm.. Without a VAE the 8Gb card has just enough room to generate 1024x1024 (or smaller) images. With a VAE, it just gets stuck likely trying to allocate VRam and is unable to, as it's not quite refusing outright.

I'm still basically using Auto1111, maybe i should familiarize myself with the other tools like ComfyUI, which I'd hoped i could avoid.

Though putting it overnight image generation I've gotten some promising images, like Tomb Raider.

 

[alebeard]

Mhmm.. Without a VAE the 8Gb card has just enough room to generate 1024x1024 (or smaller) images. With a VAE, it just gets stuck likely trying to allocate VRam and is unable to, as it's not quite refusing outright.

I'm still basically using Auto1111, maybe i should familiarize myself with the other tools like ComfyUI, which I'd hoped i could avoid.

Though putting it overnight image generation I've gotten some promising images, like Tomb Raider.

View attachment 2065086View attachment 2065087View attachment 2065089View attachment 2065088View attachment 2065093View attachment 2065094View attachment 2065095View attachment 2065098View attachment 2065092View attachment 2065096View attachment 2065097View attachment 2065099

I use the pony VAE and an 8gb card with stable difussion forge neo as the base https://github.com/Haoming02/sd-webui-forge-classic/tree/neo It's definitely not as fast as my 16 gig card on my desktop but is definitely usable

 

[yano2mch]

Got two models that are promising. Diamond XL and Hezi League of legends style.

Basically the same prompt to make Red riding hood (Rated R).

 

[fookeroo]

Help a guy out get some free referral coins.

Visit the link if you want.

1. https://imaginex.video/r/REF4HDJWD


2. https://video.a2e.ai/?coupon=Y4SI



3. unlucid.ai https://unlucid.ai/r/rgkfkt13

 

[yano2mch]

I WAS just about to recommend a model for generation, but the user and model suddenly disappeared... Very weird.

The model was Jrd Green Whisper SDXL.

edit: Found the model still accessible via the archive. So it's not lost, along with several other models i rather liked.

 

[yano2mch]

The 'babes' model looks like it can get very close to KLRXO's works, in case we wanted some covers... Probably some additional loras for specific actions we could get closer to KLRXO's style. But I'm really only considering for story covers for ones that are lacking, or if anyone wanted a KLRXO inspired cover for an unrelated story.

 

[yano2mch]

Another promising model. Unholy desires Temptation, so far every image has been really nice. Even with really vague descriptions.

 

[FreedaBees]

The image is from a book cover. The desired result would be an as near identical image generated from scratch with better resolution, no over exposure of the figure and minus the damage. The intended usage is to replace the existing image. The resolution need not be more than 800p on the vertical axis.
Thanks