AI Technical Discussion and Info

[joshbremmer2]

Gentlemen, hold on to your butts. Just figured out how to start generating videos....

awsome, what did you use ?

 

[Tezcat]

awsome, what did you use ?

futzed around on comfyworkflows until I got something that worked!

awsome, what did you use ?

 

[boobhunterTX]

civitai is setting up their stuff to a new provider, runs very very slow at the moment and they have disabled nsfw generations, .... hope they'll fix it soon.

 

[joshbremmer2]

Found this : https://github.com/cheald/sd-webui-loractl , it's a game changer for me, now able to use loras for bigger breasts only at beginning of image generation, and even if it's very manga like, it's fixed at the end of generation

 

[boobhunterTX]

interesting new feature at krea.ai: generate an image in seconds and dress up your model while growing boobs bigger same time

 

[SonicIsBack]

That'd be great. I'd even pay money to buy this model when its up and running smoothly.

It feels a bit icky selling LORAs, I've actually got 2 or 3 that I'm really happy with, I combine them to make some pretty crazy stuff.

 

[Jørgen_Brønlund]

anyone know of a good undressing app? thanks

I am starting on page one of this thread, so this may have been answered:

Nearly all of the AI apps work by looking at a series of data and then "training" on it.

The old-school "nudify" apps attempted to:

• look at a single image that you uploaded
• look at the clothed boobs that it had been trained on
• and then replace the boobs

Since the new apps are so much more sophisticated, and can work with much larger images, you might consider this:

• First, you want to find as many bikini pics as possible, ideally showing their full body
• one by one, pop them into Yandex image search, and pull up "similar" images. That will find you women who have a similar body. Once you find someone with a similar body, find as many nudes of them as you possibly can. We're basically looking for a nude model who has a similar body. You want a lot of images because the more you train, the better results you get
• Then get as many pics of the person you're trying to nudify.
• Now train your AI on the combined set of the two. The idea is to train the ai on the naked body *and* the clothed body.
• This is speculation on my part, and this sounds kinda creepy, but if you crop out the head of the nude model I believe it will force the AI to only use the face of the clothed model

All of this sounds like something out of "The Thing" but I think it should work

 

[Jørgen_Brønlund]

So, I got this error and wrestled with it for a couple days.
The solution looks like you need a degree in Python programming language.
Although, I did see comments about decreasing batch size and the number of batches, and so on.
What finally worked for me was to keep my width and height measurements reasonable. I was pushing out 800x1000 and have decreased to 500x700. It seems to have fixed it up for me.

I've been using AI to upscale my BluRay collection, so I have something like five PC workstations with various GPUs

Particularly with sophisticated AI, you need quite a bit of VRAM

The amount of RAM on the system itself doesn't play a huge factor, but if you have insufficient VRAM it will just hit a wall

If you want to find out what cards offer the best bang for the buck, there are stable diffusion benchmarks over on Reddit

IIRC, the Nvidia 3060 12GB (or is it 3060TI 12GB?) might be the 'sweet spot.' Reasonably affordable and has enough VRAM to handle most of what you throw at it

You can get used enterprise Nvidia cards on eBay with 16GB+, but I vastly prefer the consumer cards because they offer way more firepower for the money. Those used enterprise cards offer a lot of RAM, but because they're generally 4-5 years out-of-date, they're fairly slow. Even worse, they're made to fit in rackmount servers. So they are limited by the heat they generate. One of my AI cards is an Nvidia 2070 super, and it has so many fans it's like a hair dryer. Many of my AI jobs run for a day or more, so the last thing I need is to have performance throttled by too much heat, or even worse, crash the computer. Even little things like the orientation of the computer make a difference. (I keep my workstations horizontal so that the weight of the video card doesn't damage the PCIE slots.)

 

[joshbremmer2]

It feels a bit icky selling LORAs, I've actually got 2 or 3 that I'm really happy with, I combine them to make some pretty crazy stuff.

Could you share them with us ?

 

[helloworld2358]

BREAK and prompt editing are to reduce concept bleeding and color bleeding. There might be tools for that.

I used Seaart from my phone. Epicphotogasm is the most realistic model. The only breast LORA that worked with it was breast size slider. I tried JK macromastia and back breaking breasts.

Example generation:
Checkpoint: Epicphotogasm last unicorn
LORAs and strength:
skin_texture_slider 2
breast size slider .75

CFG 7, steps 40, sampler DPM++ 2M Karras, seed 1432172392

Prompt:
+
pale white (adult woman) with (pale_white_skin:1.4), [(beauty_marks:1)beauty_marks:1.4):4], [(Natural light_blonde_hair:1.1), (light_blonde_eyelashes, light_blonde_eyebrows:1.3):6], [(blue_eyes:1.2):6], (nude, naked:1.5), (medium shot:1), (full body shot:.1), Clean face, clean breasts, detailed_skin, (skinny:1.4), (fit:1.45), (sunny:1.1), front light, well lit, (natural_gigantic_breasts, huge_breasts, busty:.9), unedited photograph, detailed, happy, high resolution, photograph, RAW photo, in frame, BREAK [(large beauty_marks, brown moles, brown beauty_marks:.1) (many large moles:1.5):12],

-
[airbrushed, suntan, dark_skin_tone:3], (child:1.1), (bad anatomy:1.1), duplicate breasts, extra_breasts, extra hands, [ng_deepnegative_v1_75t::16], extra limbs, extra_nipples, long_torso, deformed areolae, deformed arms, deformed hands, [makeup:3] lacerations, acne, rosacea, vitiligo, [dirty face, shady, (mascara, eyeshadow), perfect_skin:3], anus, long_body, amputated, (piercing, jewelry, clothes, dressed, lingerie, bikini), fat, (thick_lips), open_mouth, out_of_frame, closeup, lipstick, censored, 6 fingers, cartoon, anime, 3d_render, (short hair:.7), low_resolution, [dark_brown_eyelashes, black_eyelashes, dark_brown_eyebrows, dark_brown_hair:3]

 

[SonicIsBack]

BREAK and prompt editing are to reduce concept bleeding and color bleeding. There might be tools for that.

I used Seaart from my phone. Epicphotogasm is the most realistic model. The only breast LORA that worked with it was breast size slider. I tried JK macromastia and back breaking breasts.

Example generation:
Checkpoint: Epicphotogasm last unicorn
LORAs and strength:
skin_texture_slider 2
breast size slider .75

CFG 7, steps 40, sampler DPM++ 2M Karras, seed 1432172392

Prompt:
+
pale white (adult woman) with (pale_white_skin:1.4), [(beauty_marks:1)beauty_marks:1.4):4], [(Natural light_blonde_hair:1.1), (light_blonde_eyelashes, light_blonde_eyebrows:1.3):6], [(blue_eyes:1.2):6], (nude, naked:1.5), (medium shot:1), (full body shot:.1), Clean face, clean breasts, detailed_skin, (skinny:1.4), (fit:1.45), (sunny:1.1), front light, well lit, (natural_gigantic_breasts, huge_breasts, busty:.9), unedited photograph, detailed, happy, high resolution, photograph, RAW photo, in frame, BREAK [(large beauty_marks, brown moles, brown beauty_marks:.1) (many large moles:1.5):12],

-
[airbrushed, suntan, dark_skin_tone:3], (child:1.1), (bad anatomy:1.1), duplicate breasts, extra_breasts, extra hands, [ng_deepnegative_v1_75t::16], extra limbs, extra_nipples, long_torso, deformed areolae, deformed arms, deformed hands, [makeup:3] lacerations, acne, rosacea, vitiligo, [dirty face, shady, (mascara, eyeshadow), perfect_skin:3], anus, long_body, amputated, (piercing, jewelry, clothes, dressed, lingerie, bikini), fat, (thick_lips), open_mouth, out_of_frame, closeup, lipstick, censored, 6 fingers, cartoon, anime, 3d_render, (short hair:.7), low_resolution, [dark_brown_eyelashes, black_eyelashes, dark_brown_eyebrows, dark_brown_hair:3]

Personally I prefer using less prompting and more LORAs and textual Inversions/Embeddings, tweaking the LORA weights to get the desired result. Here's an example:

Steps: 15, Sampler: Euler a, CFG scale: 1.5, Seed: 608069851, Size: 512x650, Model hash: 2fb85c113b, Model: Corbe_Realism_Mix.fp16, VAE hash: d9fcdfe0b8, VAE: vae-ft-mse-840000-ema-pruned.safetensors

Positive:
best quality, masterpiece, detailed, realistic, girl, MastasiaV2, gigantic veiny breasts, (((ginger hair))) <lora:head_slider_v2:-1><lora:breastsizeslideroffset:0.2> <lora:skin_tone_slider_v1:-1> <lora:hyperbreasts_v6:0.8> <loraAseer-SD15-LCM Quick:1> <lora:bigbigbig-08:0.4> <lora:skin_slider_v1:1> <lora:VeinyBreasts70EpochMaskedReducedMore64_1:0.5> <lora:mvtits:0.2> <lorald_siemens:0.1> <lora:skin_tone_slider_v1:-1.5> <lora:MastasiaV2-10:1>

Negative:
EasyNegative, bad_prompt_version2, multiple-breasts:1.4, multiple-nipples:1.4

Hires upscale: 2, Hires steps: 15, Hires upscaler: 4x_NMKD-Superscale-SP_178000_G

Link to my checkpoint merge model here

The key is creating your own LORAs, using good quality images. I use this guide, its Google colab, so you don't need a powerful PC/GPU.

This LORA is absolutely essential in my opinion, it cuts down generation time and gives much better detail/faces/contrast

 

[SonicIsBack]

Could you share them with us ?

If I wasn't monetizing my work I would be more than happy to share them, I have already decided that once I hit my goal of subscribers to my patreon I will be either sharing my LORAs or adding another tier to access them. I'm not in it for the money per se but times are hard, even with a full time job!

 

[Bna7926]

I have tried AI for 3 months now.

No matter how I try, AI can only create Milena Velba's boobs.

AI cannot create huge boobs pointing down, Mschocolitt boobs.

 

[Drewpournelle]

I have tried AI for 3 months now.

No matter how I try, AI can only create Milena Velba's boobs.

AI cannot create huge boobs pointing down, Mschocolitt boobs.

I went looking for a LORA but all the sagging and hanging ones still have the nipples pointing front. I've never trained a LORA but I think it's mostly just getting the pictures together and then letting the computer churn for a while.

 

[SonicIsBack]

I went looking for a LORA but all the sagging and hanging ones still have the nipples pointing front. I've never trained a LORA but I think it's mostly just getting the pictures together and then letting the computer churn for a while.

Literally this, you need to make a LORA

 

[SonicIsBack]

I have tried AI for 3 months now.

No matter how I try, AI can only create Milena Velba's boobs.

AI cannot create huge boobs pointing down, Mschocolitt boobs.

Like this?

 

[BustyLadyLover]

Like this?

If someone had tits like that, they should be smiling! ;-)

 

[Bna7926]

Like this?

Nice, this is what I want.

Do you mind PM share to me how to do it? TQ very much.

 

[SonicIsBack]

Nice, this is what I want.

Do you mind PM share to me how to do it? TQ very much.

It's a LORA I made, I don't mind sharing it once I'm back on my pc

 

[boobhunterTX]

Your work is amazing, ever thought about making video tutorials, even on a paid base?