AI Technical Discussion and Info

[SonicIsBack]

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

Me? If so, thank you . I have thought about it yeah, but due to my current circumstances it would be very difficult to do unfortunately.

 

[eric69]

I've had a read through this and I'm totally at sea.

Does anyone have anything like a ten step guide for us simple folk who are clueless about it all, but want to try their hand at making AI huge tit bimbos?

Like, Step 1, sign up to this and download that programme.
Step 2, in the first box type in,
Step 3, in the second box type in,
Step 4, wave a magic wand and utter these words....

 

[Ptar]

I've had a read through this and I'm totally at sea.

Does anyone have anything like a ten step guide for us simple folk who are clueless about it all, but want to try their hand at making AI huge tit bimbos?

Like, Step 1, sign up to this and download that programme.
Step 2, in the first box type in,
Step 3, in the second box type in,
Step 4, wave a magic wand and utter these words....

100% Minghia is the best member here, in my opinion, for using a non-website AI generator.

Did you see this post:
https://titsintops.com/phpBB2/index...cal-discussion-and-info.13545025/post-2282273

I believe there are now newer/better? updates, but this is a good place to start.

 

[eric69]

Did you see this post:
https://titsintops.com/phpBB2/index...cal-discussion-and-info.13545025/post-2282273

I believe there are now newer/better? updates, but this is a good place to start.

Yes, I looked at that post and it's a bit technical for me. I'm not bad with computer stuff, but it's all totally different to anything I've tried.

 

[Ptar]

Yes, I looked at that post and it's a bit technical for me. I'm not bad with computer stuff, but it's all totally different to anything I've tried.

It takes a bit of work, but I was able to get it rolling.
I found the tutorial well written, which helped, along with asking questions here.

 

[rgb67]

Yes, I looked at that post and it's a bit technical for me. I'm not bad with computer stuff, but it's all totally different to anything I've tried.

The first thing to figure out is if you want to run it locally or not. You need a decent graphics card to run these models on your own computer, and it takes the most work setting up, but it gives you more freedom.

There are some newer programs for running stable diffusion locally which are supposed to be easier to install, like Invoke-AI, Fooocus, and Easy Diffusion. I haven't used any of these myself.

 

[SonicIsBack]

Yes, I looked at that post and it's a bit technical for me. I'm not bad with computer stuff, but it's all totally different to anything I've tried.

Also, it's never going to be as easy as just clicking a few buttons and getting good results, you will need to do your research and put in the hours to get high quality pictures.

 

[eric69]

Also, it's never going to be as easy as just clicking a few buttons and getting good results, you will need to do your research and put in the hours to get high quality pictures.

I can believe it. It's just not having much of a clue to begin with.

 

[41boner]

I have a 3060ti, and for me SDXL takes too long. Easier to use SD 1.5 and upscale, although the negative prompts have to be on point.

EDIT: found a way to make sdxl faster.

 

[mauroao]

I continued training the Lora from the previous post and added an updated download link. The likenesses seem to be better, and I uploaded the full 300MB Lora since I'm putting it on an external host anyway.

This was my training command:

Spoiler: Full training command

python accelerate launch --num_cpu_threads_per_process=2 ./train_network.py \
--pretrained_model_name_or_path=${MODEL_DIR}/models/Stable-diffusion/v1-5-pruned.ckpt \
--train_data_dir=${TRAINING_DIR}/selected_subjects \
--output_dir=${SD_DIR}/models/Lora/trained \
--logging_dir=${SD_DIR}/models/Lora/trained/logging \
--network_weights=${SD_DIR}/models/Lora/trained/kohya_selected_subjects_v2-000020.safetensors \
--sample_prompts=user/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_resume \
--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=0.9,0.99" "d_coef=1.0" "weight_decay=0.01" "use_bias_correction=True" "safeguard_warmup=True" \
--train_batch_size=8 \
--scale_weight_norms=2 \
--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 main changes were adding the "network_weights" option to resume training, increasing the batch size, tweaking a couple of "optimizer_args" (not sure if that made any difference), and changing "scale_weight_norms" from 1 to 2. scale_weight_norms is an option that helps prevent overfitting, and changing the value from 1 to 2 makes its effect weaker.

Is this first model Sexy Venera based ?
It looks like Venera body shape and face (but without the humougous breasts....).

This is a very good job !
I want to do the same: Traine an A.I. model to reproduce a specific busty girl pics.
Maybe in the future, we will be able to create videos too !

 

[boobhunterTX]

a friend of mine just bought a HP Z8 fury workstation for his company, with 3 nvidia ada rtx 6000. Will try out a few things this weekend.

 

[rgb67]

Is this first model Sexy Venera based ?
It looks like Venera body shape and face (but without the humougous breasts....).

This is a very good job !
I want to do the same: Traine an A.I. model to reproduce a specific busty girl pics.
Maybe in the future, we will be able to create videos too !

The first image is based on her. I've found it hard to get breasts just right on SD1.5 models. I think the model merges are overfit to a generic breast shape. There might be a better way to train that avoids this. I train the Lora on top of the base model then use it with something like Realistic Vision.

With SDXL the results look better. There's less of a gap between the SDXL base model and the popular merges people use, plus the model is better in general. I haven't worked on any of this in a while, but I did make an SDXL Lora a while back similar to the SD1.5 ones:

https://www.dosyaupload.com/27nw2/subjects_xl_v1.zip

Some prompting instructions and the list of women included is in the zip. (sorry, this one doesn't include Venera)

 

[Connoisseur Junky]

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

If you are still willing to share it I would really appreciate it!

 

[SonicIsBack]

If you are still willing to share it I would really appreciate it!

Here it is!

It will need some tinkering, as all LORAs do, so try different weights, checkpoints and always use LCM with Euler a (it gives the best results for everything)

 

[Jørgen_Brønlund]

I've had a read through this and I'm totally at sea.

Does anyone have anything like a ten step guide for us simple folk who are clueless about it all, but want to try their hand at making AI huge tit bimbos?

Like, Step 1, sign up to this and download that programme.
Step 2, in the first box type in,
Step 3, in the second box type in,
Step 4, wave a magic wand and utter these words....

Here's what works for me:

step 1: go and download "automatic1111". This is a GUI for stable diffusion. Follow the installation instructions. I've installed automatic1111 so many times now, I can get it running in about 15-20 minutes now. Most of that twenty minutes is spent downloading the binaries. In order to run it locally, you'll need to be comfortable with running programs on the command line in Windows. But it's literally one command, so no need to worry.

step 2: The output of your images is going to be massively influenced by which model you use. There are dozens. I have no use for anything before SDXL, I just don't know why anyone would use two year-old technology when all of this is moving fast. For me, the "base" SDXL model probably produces the best results, but getting it to make nudes is a little tricky. You have to add prompts like "nudist" or "nude" and if you include any clothes whatsoever, it tends to clothe the model entirely. So, SDXL is good for nudes but not great for topless pics. It CAN do topless pics, but if you tell it to make an image with "a woman in a red bikini bottom" it will tend to give her a top too. The model named "Juggernaut XL" is much easier to coax nudes out of, but I find it's images to be a little rigid. Basically, the original SDXL model (the one which is theoretically "safe for work") is much more flexible, you just have to know how to trick it into making nudes.

step 3: you can add Loras for all kinds of stuff, for instance I'm about to post some info on how I made a Lora of Lucie Wilde.

 

[Jørgen_Brønlund]

OK, I made an AI of Lucie Wilde, here's an example:

https://titsintops.com/phpBB2/index.php?attachments/00055-4064878949-topaz-jpg.4598495/

For me, the big breakthrough was stumbling across some posts on red dit, from a user who came up with a fantastic configuration file for training Loras.

Here's his config file, this single file basically took my images from "sorta ok" to "freaking great"

I'm not sure if I can link to red dit, but you can google any line in this file and you should get the URL

{

"LoRA_type": "Standard",

"adaptive_noise_scale": 0.00357,

"additional_parameters": "--log_prefix=xl-lora",

"block_alphas": "",

"block_dims": "",

"block_lr_zero_threshold": "",

"bucket_no_upscale": false,

"bucket_reso_steps": 32,

"cache_latents": true,

"cache_latents_to_disk": true,

"caption_dropout_every_n_epochs": 0.0,

"caption_dropout_rate": 0,

"caption_extension": ".txt",

"clip_skip": "1",

"color_aug": false,

"conv_alpha": 4,

"conv_block_alphas": "",

"conv_block_dims": "",

"conv_dim": 4,

"decompose_both": false,

"dim_from_weights": false,

"down_lr_weight": "",

"enable_bucket": true,

"epoch": 100,

"factor": -1,

"flip_aug": false,

"full_bf16": false,

"full_fp16": false,

"gradient_accumulation_steps": 1,

"gradient_checkpointing": true,

"keep_tokens": 1,

"learning_rate": 0.0,

"logging_dir": "C:\\training\\breasts-xl\\logging",

"lora_network_weights": "",

"lr_scheduler": "cosine",

"lr_scheduler_args": "",

"lr_scheduler_num_cycles": "",

"lr_scheduler_power": "",

"lr_warmup": 10,

"max_bucket_reso": 2048,

"max_data_loader_n_workers": "0",

"max_resolution": "1024,1024",

"max_timestep": 1000,

"max_token_length": "75",

"max_train_epochs": "",

"max_train_steps": "",

"mem_eff_attn": false,

"mid_lr_weight": "",

"min_bucket_reso": 256,

"min_snr_gamma": 5,

"min_timestep": 0,

"mixed_precision": "bf16",

"model_list": "custom",

"module_dropout": 0,

"multires_noise_discount": 0,

"multires_noise_iterations": 0,

"network_alpha": 128,

"network_dim": 128,

"network_dropout": 0.1,

"no_token_padding": false,

"noise_offset": 0.0357,

"noise_offset_type": "Original",

"num_cpu_threads_per_process": 2,

"optimizer": "Prodigy",

"optimizer_args": "weight_decay=0.05 betas=0.9,0.98",

"output_dir": "C:\\training\\breasts-xl\\lora-output",

"output_name": "breasts-sdxl",

"persistent_data_loader_workers": false,

"pretrained_model_name_or_path": "C:\\github\\stable-diffusion-webui\\models\\Stable-diffusion\\sdxl\\sd_xl_base_1.0_0.9vae.safetensors",

"prior_loss_weight": 1.0,

"random_crop": false,

"rank_dropout": 0,

"reg_data_dir": "",

"resume": "",

"sample_every_n_epochs": 0,

"sample_every_n_steps": 50,

"sample_prompts": "photo of a woman, smile, blonde hair, medium breasts, nipples --w 832 --h 1248",

"sample_sampler": "euler_a",

"save_every_n_epochs": 10,

"save_every_n_steps": 0,

"save_last_n_steps": 0,

"save_last_n_steps_state": 0,

"save_model_as": "safetensors",

"save_precision": "bf16",

"save_state": false,

"scale_v_pred_loss_like_noise_pred": false,

"scale_weight_norms": 1,

"sdxl": true,

"sdxl_cache_text_encoder_outputs": true,

"sdxl_no_half_vae": true,

"seed": "",

"shuffle_caption": false,

"stop_text_encoder_training": 0,

"text_encoder_lr": 0.0,

"train_batch_size": 4,

"train_data_dir": "C:\\training\\breasts-xl\\img",

"train_on_input": false,

"training_comment": "",

"unet_lr": 1.0,

"unit": 1,

"up_lr_weight": "",

"use_cp": false,

"use_wandb": false,

"v2": false,

"v_parameterization": false,

"v_pred_like_loss": 0,

"vae_batch_size": 0,

"wandb_api_key": "",

"weighted_captions": false,

"xformers": "xformers"

}

 

[Connoisseur Junky]

Here it is!

It will need some tinkering, as all LORAs do, so try different weights, checkpoints and always use LCM with Euler a (it gives the best results for everything)

Thanks for sharing!.. i'll give it a try does it have any keywords?

btw - what do you mean LCM with Euler.. I thought LCM is a sampler?

do you use it in the workflow like in this example? https://www.youtube.com/watch?v=ZzjSjwKrjUE

 

[SonicIsBack]

Thanks for sharing!.. i'll give it a try does it have any keywords?

btw - what do you mean LCM with Euler.. I thought LCM is a sampler?

do you use it in the workflow like in this example? https://www.youtube.com/watch?v=ZzjSjwKrjUE

Keyword is BMButtons (if you keep the weight at 1 you will get pictures from the neck down so try 0.4 to 0.6) , LCM is a LORA and a sampler but I don't use the LCM sampler as it doesnt give the best details, the PAseer version is the best one. I usually use 1.5 or 2 cfg, and 12-15 steps with Euler a, keep the LORA weight at 1.

 

[rgb67]

Keyword is BMButtons (if you keep the weight at 1 you will get pictures from the neck down so try 0.4 to 0.6) , LCM is a LORA and a sampler but I don't use the LCM sampler as it doesnt give the best details, the PAseer version is the best one. I usually use 1.5 or 2 cfg, and 12-15 steps with Euler a, keep the LORA weight at 1.

Thanks for sharing that Lora. I made a few pics with it here.

I'm getting back to training now, and it's been a while. Has anyone used OneTrainer instead of Kohya, or tried using LLaVA for automatic captioning?

 

[SonicIsBack]

Thanks for sharing that Lora. I made a few pics with it here.

I'm getting back to training now, and it's been a while. Has anyone used OneTrainer instead of Kohya, or tried using LLaVA for automatic captioning?

Nice results! I use the google colab by HollowStrawberry for all my LORA training, there is an SDXL version now too.