AI Technical Discussion and Info

[Orhm]

Is there a website where you can upload a picture and have it be AI generated? I used one before but for the life of me cannot remember it.
I used 2 pictures of Angelica Stranger. but I honestly have no idea what I selected and which format? but all I know is they don't look like her, lol.

 

[goodsexboy69]

Is there a website where you can upload a picture and have it be AI generated? I used one before but for the life of me cannot remember it.
I used 2 pictures of Angelica Stranger. but I honestly have no idea what I selected and which format? but all I know is they don't look like her, lol.

What program did you use to take these images?

 

[Ptar]

What program did you use to take these images?

I believe he said he can't remember it.


https://titsintops.com/phpBB2/index...cal-discussion-and-info.13545025/post-2282273
You should check out Stable Diffusion.

 

[maxiquick]

Here's some info about Lora training that I've gathered, plus a Lora I just made on pics of 106 models and random celebrities. (all names are in the zip file, and I avoided DNP models)

Couldn't add the Lora as an attachment for some reason: https://www.dosyaupload.com/27cog/selected_subjects_v2.zip
Edit: New version here: https://www.dosyaupload.com/1rf33/selected_subjects_v3.zip

There's many full guides out there with more information. I'd recommend "The Other LoRA Training Rentry", which has lots of detailed info and is up-to-date.

My strategy is to take my folders of images of different women, tag them with the WD14 tagger, and discard images with certain tags (like "2girls" to filter out group pictures). Then I randomly take some number from each, 50 in this case, and train on them. The images are captioned with the full name plus the WD14 tags. For some women, I manually added "breast implants" and "butterface" as tags. The "breast implants" tag works as expected and is great as a negative prompt, but the "butterface" tag didn't end up capturing what I wanted it to.

The tool I use for training is Kohya's SD scripts. Some people have made easier interfaces for this tool, like bmaltais is probably the most popular one and has WD14 tagging built in.

Once I have my training data prepared with a mess of Python scripts I wrote, I ran training with this command in Linux. You'd have to adapt it for Windows and fill in your own directories. I'm mainly putting it here for reference:

Spoiler: Full train_network.py command

python accelerate launch --num_cpu_threads_per_process=2 ./train_network.py \
--pretrained_model_name_or_path=${MODEL_DIR}/v1-5-pruned.ckpt \
--train_data_dir=${TRAINING_DIR}/selected_subjects \
--output_dir=${SD_DIR}/models/Lora/trained \
--logging_dir=${SD_DIR}/models/Lora/trained/logging \
--sample_prompts=cleaned_subjects_prompts.txt \
--enable_bucket --resolution=512,512 --bucket_no_upscale \
--network_alpha=128 --network_dim=256 \
--save_model_as=safetensors --network_module=networks.lora \
--output_name=kohya_selected_subjects_v2 \
--sample_every_n_steps=10000 \
--no_half_vae \
--learning_rate=1 --lr_scheduler=constant --lr_warmup=0 --optimizer_type="Prodigy" \
--optimizer_args "decouple=True" "d0=1e-5" "betas=0.9,0.999" "d_coef=1.0" "weight_decay=0.01" "use_bias_correction=False" "safeguard_warmup=False" \
--train_batch_size=2 \
--scale_weight_norms=1 \
--random_crop --caption_tag_dropout_rate=0.05 \
--save_every_n_epochs=5 --max_train_epochs=50 \
--mixed_precision=bf16 --save_precision=bf16 --seed=12345 --caption_extension=.txt --max_data_loader_n_workers=0 \
--bucket_reso_steps=32 --xformers

The model that I train on is the SD1.5 base model.

I'm running on a 24GB GPU, and you may need to adjust some of these settings to get training to fit into memory. Lowering the batch size and adding --gradient_checkpointing should help.

I set 50 epochs for training, but didn't really want to train that long. After training was done with 20 epochs, I killed the training job and resized the Lora with this command:

Spoiler: Resize Lora command

python networks/resize_lora.py --save_precision=fp16 --new_rank=24 \
--model=${SD_DIR}/models/Lora/trained/kohya_selected_subjects_v2-000020.safetensors
--save_to=${SD_DIR}/models/Lora/trained/kohya_selected_subjects_v2_e20_resize24.safetensors

The likenesses aren't perfect here, and there's a lot I could improve in my training. Better image filtering (removing low-quality and messy images) and better captioning are the first things I'd want to improve. Messing around with training settings could also help, and if you have suggestions of things to change let me know. I understand some of these settings well, some a little bit, and some not at all. Feel free to ask about them.

Hi, man
Your work with your LORA is excellent. It's really good to see.
I'm starting to play around with creating photos using Tensor.art since I don't have a GPU. I'm trying to learn the use of prompts, models, controlnet, etc.
I would really like to use your LORA to continue learning and have fun at the same time.
Is there any chance you can upload it to civitai? I think from there I can import it to Tensor.art

I promise to share the photos I can create using your LORA. I have no problem sharing the promps I'm using.

I share with you some recent creations as a sign of good faith.

 

[rgb67]

Hi, man
Your work with your LORA is excellent. It's really good to see.
I'm starting to play around with creating photos using Tensor.art since I don't have a GPU. I'm trying to learn the use of prompts, models, controlnet, etc.
I would really like to use your LORA to continue learning and have fun at the same time.
Is there any chance you can upload it to civitai? I think from there I can import it to Tensor.art

I promise to share the photos I can create using your LORA. I have no problem sharing the promps I'm using.

I share with you some recent creations as a sign of good faith.

I've thought about uploading something to Civitai but haven't got around to it. I don't mind if you upload it yourself to use it, either to civitai or tensor.art. Just mention someone else trained it and add a link to my post if you can.

For Civitai, they have a policy that for models that reproduce a person's likeness you can't have any lewd preview images. So even for porn stars, you need to make sure they're covered up.

 

[maxiquick]

I've thought about uploading something to Civitai but haven't got around to it. I don't mind if you upload it yourself to use it, either to civitai or tensor.art. Just mention someone else trained it and add a link to my post if you can.

For Civitai, they have a policy that for models that reproduce a person's likeness you can't have any lewd preview images. So even for porn stars, you need to make sure they're covered up.

Thank you for your permission to use your LORA!
I'll try to upload it to Tensor.art, I'm very inexperienced with these things, but if I manage to do it I'll let you know. Anyway, in that case you will see the photos here.
Please detail here exactly the link you want me to refer to if I am successful with this task.
From what I have seen, since the lora is based on pornstars, the images created cannot be used commercially, but you can upload and use the lora by notifying others of the conditions.

 

[rgb67]

Thank you for your permission to use your LORA!
I'll try to upload it to Tensor.art, I'm very inexperienced with these things, but if I manage to do it I'll let you know. Anyway, in that case you will see the photos here.
Please detail here exactly the link you want me to refer to if I am successful with this task.
From what I have seen, since the lora is based on pornstars, the images created cannot be used commercially, but you can upload and use the lora by notifying others of the conditions.

You can link to this post:

https://titsintops.com/phpBB2/index.php?goto/post&id=2319408

 

[maxiquick]

You can link to this post:

https://titsintops.com/phpBB2/index.php?goto/post&id=2319408

hey man, I was able to upload your Lora, both to Civitai and Tensor.art
I don't know if I have done it correctly, the links are the following:
https://civitai.com/models/184437?modelVersionId=207030
https://tensor.art/models/654990024808559584
I haven't put much energy into the descriptions due to lack of time.

I have tried to create an image but it hasn't turned out very well, I will try again at another time.
It would be good for me if you could tell me what base model you use to create the images so that they look as similar as possible to the Lora models.

 

[rgb67]

hey man, I was able to upload your Lora, both to Civitai and Tensor.art
I don't know if I have done it correctly, the links are the following:
https://civitai.com/models/184437?modelVersionId=207030
https://tensor.art/models/654990024808559584
I haven't put much energy into the descriptions due to lack of time.

I have tried to create an image but it hasn't turned out very well, I will try again at another time.
It would be good for me if you could tell me what base model you use to create the images so that they look as similar as possible to the Lora models.

The model I'm using is Realistic Vision, and the only other resource I'm using is one negative embedding. With this Lora, I've found that I need to set the CFG pretty low (3.0) to get good results.

Here are some images with the generation info from SD.Next stored in them. I'll also put the full generation info here:

Spoiler: Full generation info

Petra Verkaik dressed as a nun, habit, smile, lips, looking at viewer, realistic, black gown, cross necklace
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK sexy, breasts, cleavage, breast implants, text, watermark, nude, topless, nipples
Steps: 20, Seed: 3511022521, Sampler: Euler a, CFG scale: 3, Size: 400x600, Batch: 1x8, Index: 1x7, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 20, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 600x900, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

Gianna Michaels dressed as a nun, habit, smile, lips, looking at viewer, realistic, black gown, cross necklace
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK sexy, breasts, cleavage, breast implants, text, watermark, nude, topless, nipples
Steps: 20, Seed: 2341914865, Sampler: Euler a, CFG scale: 3, Size: 400x600, Batch: 1x8, Index: 1x4, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 20, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 600x900, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

Danica Collins, topless, breasts, nipples, panties, cowboy shot, realistic, smirk, dark skin, dark-skinned female
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK breast implants, text, watermark
Steps: 20, Seed: 3550037121, Sampler: Euler a, CFG scale: 3, Size: 400x600, Batch: 1x8, Index: 1x8, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 20, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 600x900, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

Danica Collins, Astronaut outfit, lips, looking at viewer, realistic, space
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK breast implants, text, watermark, nude, topless, nipples
Steps: 20, Seed: 350411048, Sampler: Euler a, CFG scale: 3, Size: 400x600, Batch: 1x8, Index: 1x1, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 30, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 600x900, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

Milena Velba, breasts, turtleneck, long hair, lips, looking at viewer, grey sweater, sitting realistic, jeans
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK breast implants, text, watermark, nude, topless, nipples
Steps: 20, Seed: 1490388903, Sampler: Euler a, CFG scale: 3, Size: 400x600, Batch: 1x8, Index: 1x7, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 30, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 600x900, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

Lucie Wilde, breasts, turtleneck, long hair, lips, looking at viewer, red sweater, arms at side, realistic, jeans
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK breast implants, text, watermark, nude, topless, nipples
Steps: 20, Seed: 146652706, Sampler: Euler a, CFG scale: 3, Size: 360x640, Batch: 1x8, Index: 1x3, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 30, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 540x960, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

Lucie Wilde, breasts, skindentation, smile, teeth, strapless corset, panties, cleavage, long hair, lips, looking at viewer, lace, arms at side
<lora:selected_subjects_v3:1.0>
Negative prompt: realisticvision-negative-embedding BREAK breast implants, text, watermark, nude, topless, nipples
Steps: 20, Seed: 788059300, Sampler: Euler a, CFG scale: 3, Size: 360x640, Batch: 1x8, Index: 1x5, Parser: Full parser, Model: realisticVisionV51_v51VAE, Model hash: 15012c538f, Backend: Original, App: SD.Next, Version: 2336ffc, Operations: txt2img; hires, Second pass: True, Hires force: False, Hires steps: 30, Hires upscaler: Latent, Hires upscale: 1.5, Hires resize: 0x0, Hires size: 540x960, Denoising strength: 0.5, Latent sampler: DPM++ 2M SDE, Image CFG scale: 3, Lora hashes: "selected_subjects_v3: e679f8803343"

I don't know what the generation interface looks like for these websites, but there should be a way to enter in most of this info somewhere. Let me know if you have any trouble.

 

[uht]

I installed Stable Diffusion and have played with it a bit. Any tips on using it to alter existing pics?

I use this step by step guide https://average-degen.github.io/nudify.html and get pretty good results as a beginner.

 

[SonicIsBack]

I'm working on the ultimate titsintops LORA, initial results are more than promising! I find that most AI generated stuff with really huge boobs ends up looking like big implants or weirdly proportioned. I wanted something that shows a bra underneath struggling to contain them and really happy with the first test run . Once I fine tune it (with more created images) it will be epic.

 

[boobhunterTX]

Hello everybody,

tried to install SU on an 2019 Mac Pro, maxed out with a powerful GPU (AMD Radeon PRO W6800X Duo 32 GB).
Tried several things and workarrounds, watched tons of tutorials but cant get it running.
This might be a GPU issue (b/c no nvidia, no cuda) they said...

Any ideas?
Thanks.

 

[joshbremmer2]

I'm working on the ultimate titsintops LORA, initial results are more than promising! I find that most AI generated stuff with really huge boobs ends up looking like big implants or weirdly proportioned. I wanted something that shows a bra underneath struggling to contain them and really happy with the first test run . Once I fine tune it (with more created images) it will be epic.

awsome, let us know when it'q ready !

 

[Upscaler]

Hello everybody,

tried to install SU on an 2019 Mac Pro, maxed out with a powerful GPU (AMD Radeon PRO W6800X Duo 32 GB).
Tried several things and workarrounds, watched tons of tutorials but cant get it running.
This might be a GPU issue (b/c no nvidia, no cuda) they said...

Any ideas?
Thanks.

You tried SHARK? It's currently the fastest implementation of Stable Diffusion that supports AMD gpus. NMKD also supports AMD. Neither of them are as robust or as updated as often as Automatic1111 though.

But yeah you'd save yourself a lot of headaches if you had an Nvidia gpu.

 

[boobhunterTX]

You tried SHARK? It's currently the fastest implementation of Stable Diffusion that supports AMD gpus. NMKD also supports AMD. Neither of them are as robust or as updated as often as Automatic1111 though.

But yeah you'd save yourself a lot of headaches if you had an Nvidia gpu.

Thanks, I'll give it a try.

 

[Upscaler]

Here's a video walking you through installing the Automatic1111 version step by step. Once you got it installed here's another video on how to use it.

Only after you get used to using the program should you attempt to start training your own custom models using lora and dreambooth. Here's a good beginner one with loras. And finally if you want the most accurate images possible you can use this tutorial to make your own dreambooth models. Be warned though you're going to have to use command line arguments as the GUI is currently bugged.

 

[Tezcat]

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

 

[midge6682]

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

now you just need to figure how you create a video of them being bent over and fucked

 

[boobhunterTX]

now you just need to figure how you create a video of them being bent over and fucked

StableDiffusion 3.0 like: "Hold my beer ..."

 

[boobhunterTX]

I'm working on the ultimate titsintops LORA, initial results are more than promising! I find that most AI generated stuff with really huge boobs ends up looking like big implants or weirdly proportioned. I wanted something that shows a bra underneath struggling to contain them and really happy with the first test run . Once I fine tune it (with more created images) it will be epic.

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