Ver Anima V1
Trained base on Anima V0.1 base with LoRA_Easy_Training_Scripts. See about this version for training details.
Ver NOOB V-PRED V3 EQ
Trained with improved parameters based on NOOB V-PRED EQ VAE exprimental version: https://civarchive.com/models/1858821/noobai-v-pred-10-with-eq-vae
Updated Ver NOOB V-PRED V2
Trained based on NOOB V-PRED 1.0 with OneTrainer. See about this version for training details.
It can do better wrinkles on pants and glans shape in this version. Sometimes it has text around but easily to remove with negative prompts.
The adaptability to art style has weakened in this version. In this case, upweight the style lora or doing a few more trials can be helpful.
Updated Ver NOOB V-PRED V1
Trained based on NOOB V-PRED 1.0. Updated dataset to achieve better performance. It enhanced this concept on NOOB series models. See the examples first.
Only one trigger word in this version.
Updated
Used PonyXL to train this one. Updated dataset to achieve better performance. It enhanced this concept on PonyXL and related models. See the examples first.
Seperated into two concepts: bulge down leg with pants/bulge down leg with dress.
Introduction
An experimental LyCORIS, used to generate futanari with bulge down leg.
Still a lot of room for improvement.
Pros
Can generate futanari with bulge down leg.
Cons
Using 3D materials for training, the art style may become strange under high weights.
Penis twisted with legs.
Confused the glan with knees.
Bad penis shape.
A lot of trials needed.
How to use it
I recommand using it in weight 0.7-0.8.
Training details
Traingset is of about 34 images. Half of them are chosen as regularization images. Total steps are about 1500.
resolution=768*768
batch_size=2
epoch=10
network_dim=16
network_alpha=16
conv_dim=4
conv_alpha=4
clip_skip=2
Using Prodigy as optimizer, setting as the author recommend.
Description
[[subsets]]
caption_dropout_rate = 0.05
caption_extension = ".txt"
caption_tag_dropout_rate = 0.05
image_dir = "G:/dataset/concept/bulge_down_leg_V2/5_bdl"
keep_tokens = 1
name = "5"
num_repeats = 5
random_crop_padding_percent = 0.05
shuffle_caption = true
[train_mode]
train_mode = "lora"
[general_args.args]
persistent_data_loader_workers = true
vae_batch_size = 5
pretrained_model_name_or_path = ""
mixed_precision = "bf16"
gradient_checkpointing = true
gradient_accumulation_steps = 1
seed = 42
max_data_loader_n_workers = 1
max_token_length = 225
prior_loss_weight = 1.0
sdpa = true
max_train_epochs = 20
cache_latents = true
cache_latents_to_disk = true
[general_args.dataset_args]
resolution = 1024
batch_size = 5
[network_args.args]
network_dim = 32
network_alpha = 16.0
min_timestep = 0
max_timestep = 1000
network_train_unet_only = true
[optimizer_args.args]
optimizer_type = "ProdigyPlusScheduleFree"
lr_scheduler = "constant"
loss_type = "l2"
learning_rate = 1.0
unet_lr = 1.0
max_grad_norm = 1.0
[saving_args.args]
output_dir = "G:/LoRA_Easy_Training_Scripts/output"
output_name = "bulge_down_leg-anima1.0-V1-loha-dim32conv16AR0.5-NewTag"
save_precision = "bf16"
save_model_as = "safetensors"
save_every_n_epochs = 1
save_toml = true
save_toml_location = "G:/LoRA_Easy_Training_Scripts/output"
[logging_args.args]
log_prefix_mode = "disabled"
run_name_mode = "default"
[anima_args.args]
pretrained_model_name_or_path = "G:/sd-webui-forge-neo/models/Stable-diffusion/ANIMA/anima_baseV10.safetensors"
qwen3 = "G:/sd-webui-forge-neo/models/text_encoder/qwen_3_06b_base.safetensors"
vae = "G:/sd-webui-forge-neo/models/VAE/qwen_image_vae.safetensors"
qwen3_max_token_length = 512
t5_max_token_length = 512
timestep_sampling = "sigmoid"
sigmoid_scale = 1.0
[edm_loss_args.args]
edm2_loss_weighting = false
[bucket_args.dataset_args]
enable_bucket = true
min_bucket_reso = 256
max_bucket_reso = 2048
bucket_reso_steps = 64
[network_args.args.network_args]
conv_dim = 16
conv_alpha = 8.0
algo = "loha"
train_llm_adapter = "false"
[optimizer_args.args.optimizer_args]
betas = "0.95,0.99"
beta3 = "None"
weight_decay = "0"
weight_decay_by_lr = "True"
d0 = "2e-5"
d_coef = "1"
d_limiter = "True"
schedulefree_c = "0"
eps = "None"
split_groups = "True"
split_groups_mean = "False"
factored = "True"
factored_fp32 = "True"
use_bias_correction = "True"
use_stableadamw = "True"
use_schedulefree = "True"
use_speed = "False"
stochastic_rounding = "True"
fused_back_pass = "False"
use_cautious = "False"
use_grams = "False"
use_adopt = "False"
use_orthograd = "False"
use_focus = "False"