Nanakusa Nazuna - Yofukashi no uta [Pony][DoRA]

This is a DoRA version of Nanakusa Nazuna - Yofukashi no uta XL (Pony version) uploaded as a separate model to conform to the Civitai model Type field.

This is part of a small network test training comparison (LoRA, DoRA, LoKR), can reference this post for their comparisons - I can upload the LoKR version on request but I felt the LoRA/DoRA versions were better.

I'm also trying to train GLoRA with kohya_ss without much luck yet (black outputs).

Recommended prompt:

Images were manually rated with quality/aesthetic tags:

masterpiece, best quality, low quality, very aesthetic, aesthetic

Positive prompt:

nanakusa nazuna \(yofukashi no uta\), 1girl, solo, blue eyes, ringed eyes, vampire, braided hair rings, blunt bangs, short hair,
{{outfit}},
score_9, score_8_up, score_7_up, score_6_up, masterpiece, best quality, very aesthetic

Negative prompt:

(worst quality, low quality:1.1), score_4, score_3, score_2, score_1, error, bad anatomy, bad hands, watermark, ugly, distorted, signature

License

Fair AI Public License 1.0-SD

Description

This is a DoRA, not your typically made LoRA. A1111 v1.9.0 and up is required to use this. Forge users must download two files, which can be found here. Go to stable-diffusion-webui\ldm_patched\modules and replace the files there.

Kohya training config (toml):

bucket_no_upscale = true
bucket_reso_steps = 256
cache_latents = true
cache_latents_to_disk = true
caption_extension = ".txt"
clip_skip = 2
dynamo_backend = "no"
enable_bucket = true
epoch = 20
full_bf16 = true
gradient_accumulation_steps = 1
gradient_checkpointing = true
huber_c = 0.1
huber_schedule = "snr"
learning_rate = 1.0
logging_dir = "E:/work/LoRa_work/logging"
loss_type = "l2"
lr_scheduler = "cosine"
lr_scheduler_args = []
lr_scheduler_num_cycles = 1
lr_scheduler_power = 1
max_bucket_reso = 2048
max_data_loader_n_workers = 0
max_grad_norm = 1
max_timestep = 1000
max_token_length = 225
max_train_epochs = 20
max_train_steps = 153
metadata_author = "motimalu"
min_bucket_reso = 512
min_snr_gamma = 5
mixed_precision = "bf16"
multires_noise_discount = 0.3
multires_noise_iterations = 6
network_alpha = 32
network_args = [ "preset=full", "conv_dim=1", "conv_alpha=1", "rank_dropout=0",
"bypass_mode=False", "dora_wd=True", "module_dropout=0", "use_tucker=False",
"use_scalar=False", "rank_dropout_scale=False", "algo=locon", "train_norm=False",]
network_dim = 32
network_module = "lycoris.kohya"
no_half_vae = true
noise_offset_type = "Multires"
optimizer_args = []
optimizer_type = "Prodigy"
output_dir = "D:/model_output"
output_name = "pony_nanakusa_nazuna_dora_v3"
pretrained_model_name_or_path =
"C:/workspace/webui_forge_cu121_torch21/webui/models/Stable-diffusion/ponyDiffusionV6XL_v6Start
WithThisOne.safetensors"
prior_loss_weight = 1
resolution = "1024,1024"
sample_every_n_epochs = 10
sample_prompts = "D:/model_output\\prompt.txt"
sample_sampler = "euler_a"
save_every_n_epochs = 20
save_model_as = "safetensors"
save_precision = "bf16"
seed = 339491249
text_encoder_lr = 1
train_batch_size = 8
train_data_dir = "C:/data/nanakusa_nazuna"
unet_lr = 1
xformers = true