->NoobAI V-Pred | PonyDiffusion / SD 1.5
Please, check the description, there is a little bit of useful information.
Should also work for Epsilon-pred version of NoobAI + IllustriousXl, but worse.
This is a renewed version of Zheng's style LoRA. I needed to start it from scratch, because i've lost most of my datasets since PonyDiffusion times. So, it would take some time to bring it back to better quality and concept knowledge. Anyway, i hope you'll like this LoRA as well.
Inеtroduction:
This LoRA is supposed to represent older (detailed) Zheng's artstyle.
Please, be sure to use the same positive and negative quality prompt for better result.
Compatibility:
Humans: great
Anthros: great
Ponies: great
Multiple characters: great
Character LoRAs; great
Unique characters: great
Recommendations:
1.0 emphasis
Trigger (v3+):
zhengClip Skip 1
CFG 5
NoobAI 1.0 V-Pred for generations
masterpiece, best quality, highres, sweat, blushin positivesworst aesthetic, worst quality, low quality, bad quality, lowresin negatives+
bright, colorful, saturated, red theme, blue themefor CIvitAI generationsCFG Rescale 0.3 - 0.4 (if you're using v-pred version)
Use anything else in negatives only if you want to avoid or remove something from the image. For example, you don't want to see doors - "door" in negative. Style will look much better without garbage in negative prompt. Just try it.
Characters:
kamiya midori, brown hair, ponytail, ahoge, hair between eyes, sidelocks, green eyes
nanami mizuki, brown hair, long hair, twintails, blunt bangs, sidelocks, brown eyes
opera brest, blonde hair, long hair, ponytail, hair between eyes, sidelocks, green eyes, dark skin
morimoto chio, black hair, medium hair, sidelocks, purple eyes
morimoto chisato, black hair, long hair, sidelocks, purple eyes
canavalia, white hair, medium hair, sidelocks, purple eyes
lizlyre, white hair, long hair, single braid, blunt bangs, sidelocks, purple eyes
sophia peronica, purple hair, medium hair, yellow eyes
tzuyin, purple hair, long hair, blunt bangs, sidelocks, purple hairband, hair ribbon, purple ribbon, purple eyes
emilia viola, purple hair, long hair, twintails, hair between eyes, sidelocks, hair ribbon, black ribbon, purple eyes
green berennice, green hair, long hair, drill hair, drill sidelocks, purple eyes
CivitAI Link: CHARACTER EXAMPLES
Some characters has very few images of them from different angles, so, Green Berennice, Sophia Peronica or Canavalia may work slightly worse than other characters. Style may be different for newer characters like Tzuyin.
Interesting concepts:
sagging testicles - for more prominent testicles sag
stable
pulsating cumshot, urethral bulge - for visible bulging cum in urethra
somehow stable
circumcised - foreskin "line" is far away from the penis tip
previously known as "pulled foreskin" (pre-v4)
unstable (+no examples yet)
Credits:
Trained on NoobAI 0.75, Dataset of 500 pictures (v1), 40 epochs, Clip Skip - 1
Trained on NoobAI 1.0, Dataset of 650 pictures (v2), 40 epochs, Clip Skip - 1
Trained on NoobAI 1.0, Dataset of 750 pictures (v3), 40 epochs, Clip Skip - 1
Trained on NoobAI 1.0, Dataset of 810 pictures (v4), 40 epochs, Clip Skip - 1
Really big thanks to:
Zheng (Allurmilk) for beautiful artwork
Some good folks from Discord server
You, for reading, downloading, liking and reviewing this LoRA
Description
What's new?
Expanded dataset up to 810 images
Better frequent character consistency
More stable and consistent style
Better eyes quality
Thicker penises
Fixed horse penis sheath color
Characters:
Kamiya Midori
Morimoto Chio
Morimoto Chistato (Morimoto Chio's mother)
Nanami Mizuki
Opera Brest <- BETTER QUALITY
Canavalia <- BETTER QUALITY
Lizlyre <- BETTER QUALITY
Sophia Peronica <- BETTER QUALITY
Tzuyin Hsieh
Emilia Viola
Green Berennice <- BETTER QUALITY
FAQ
Comments (6)
What's new?
- Expanded dataset up to 810 images
- Better frequent character consistency
- More stable and consistent style
- Better eyes quality
- Thicker penises
- Fixed horse penis sheath color
Characters:
- Kamiya Midori
- Morimoto Chio
- Morimoto Chistato (Morimoto Chio's mother)
- Nanami Mizuki
- Opera Brest <- BETTER QUALITY
- Canavalia <- BETTER QUALITY
- Lizlyre <- BETTER QUALITY
- Sophia Peronica <- BETTER QUALITY
- Tzuyin Hsieh
- Emilia Viola
- Green Berennice <- BETTER QUALITY
do you mind sharing config/toml for training lora? its fine if not, i understand
Hi, no, it's not a secret.
But my .toml file is made for LoRA Easy Training Script and probably is incompatible with Kohya SS or anything else. Subset parameters are unique for this LoRA, so, i don't recommend to use it. Also, this LoRA was trained for V-Pred model, so, it has:
v_parameterization = true
scale_v_pred_loss_like_noise_pred = true
Anyway. There are training params:
[[subsets]]
caption_extension = ".txt"
caption_tag_dropout_rate = 0.1
image_dir = "D:/Old/Training Data/Styles/Zheng/Zheng - Allurmilk - Female"
name = "Zheng - Allurmilk - Female"
num_repeats = 1
shuffle_caption = true
[[subsets]]
caption_extension = ".txt"
caption_tag_dropout_rate = 0.1
image_dir = "D:/Old/Training Data/Styles/Zheng/Zheng - Allurmilk - Futanari"
keep_tokens = 1
name = "Zheng - Allurmilk - Futanari"
num_repeats = 1
shuffle_caption = true
[[subsets]]
caption_extension = ".txt"
caption_tag_dropout_rate = 0.1
image_dir = "D:/Old/Training Data/Styles/Zheng/Zheng - Old - Female"
name = "Zheng - Old - Female"
num_repeats = 1
shuffle_caption = true
[[subsets]]
caption_extension = ".txt"
caption_tag_dropout_rate = 0.1
image_dir = "D:/Old/Training Data/Styles/Zheng/Zheng - Old - Futanari"
keep_tokens = 1
name = "Zheng - Old - Futanari"
num_repeats = 1
shuffle_caption = true
[[subsets]]
caption_extension = ".txt"
caption_tag_dropout_rate = 0.1
image_dir = "D:/Old/Training Data/Styles/Zheng/Zheng - Original - Female"
keep_tokens = 1
name = "Zheng - Original - Female"
num_repeats = 1
shuffle_caption = true
[[subsets]]
caption_extension = ".txt"
caption_tag_dropout_rate = 0.1
image_dir = "D:/Old/Training Data/Styles/Zheng/Zheng - Original - Futanari"
keep_tokens = 2
name = "Zheng - Original - Futanari"
num_repeats = 1
shuffle_caption = true
[train_mode]
train_mode = "lora"
[general_args.args]
max_data_loader_n_workers = 1
persistent_data_loader_workers = true
pretrained_model_name_or_path = "C:/Programs/Stable Diffusion - Forge/models/Stable-diffusion/noobaiXLNAIXL_vPred10Version.safetensors"
no_half_vae = true
mixed_precision = "bf16"
gradient_checkpointing = true
seed = 23
max_token_length = 225
prior_loss_weight = 1.0
xformers = true
max_train_epochs = 40
cache_latents = true
sdxl = true
v_parameterization = true
scale_v_pred_loss_like_noise_pred = true
[general_args.dataset_args]
resolution = 1024
batch_size = 1
[network_args.args]
network_dim = 64
network_alpha = 32.0
min_timestep = 0
max_timestep = 1000
[optimizer_args.args]
optimizer_type = "AdamW8bit"
lr_scheduler_num_cycles = 2
lr_scheduler = "cosine_with_restarts"
loss_type = "l2"
learning_rate = 0.00012
warmup_ratio = 0.1
unet_lr = 0.00012
text_encoder_lr = 6e-5
max_grad_norm = 1.0
min_snr_gamma = 5
zero_terminal_snr = true
[saving_args.args]
output_dir = "C:/Users/Andrew/Desktop/Training Output"
output_name = "Zheng v4 - NoobAI 1.0 V-Pred"
save_precision = "fp16"
save_model_as = "safetensors"
save_every_n_epochs = 10
save_state = true
save_last_n_epochs_state = 1
[bucket_args.dataset_args]
enable_bucket = true
min_bucket_reso = 640
max_bucket_reso = 1024
bucket_reso_steps = 64
@TroubleDarkness thanks a bunch! do you know how much vram / ram this can fit into? mostly i am trying to train using colab
@low_channel_1503 i'm using RTX 3060 for all my training sessions, and it uses ~10 gigabytes of VRAM for training with batch size 1, and somehere between 11 and 12 with batch size of 2. Dataset size is 810 images. Optimizer - AdamW8bit. You'll probably get different VRAM usage on different settings. Don't forget to use gradient checkpointing, if your training colab requires to specify it, for some reason.
@TroubleDarkness thank you again.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.