Aesthetic Quality Modifiers - Masterpiece

Training data is a subset of all my manually rated datasets with the quality/aesthetic modifiers, including only the masterpiece tagged images.

Subset in the Aesthetic Quality Modifiers Collection.

ℹ️ LoRA work best when applied to the base models on which they are trained. Please read the About This Version on the appropriate base models, trigger usage, and workflow/training information.

Generation Settings:

Previews are generated in Forge with upscaling and adetailer.

For Noobai V-Pred, a ComfyUI workflow reference with DynamicThresholding, Upscaling, and FaceDetailer can be found here: https://civarchive.com/posts/11457095

Description

Trained on Anima Preview 2

Assume that any lora trained on the preview version won't work well on the final version.

Recommended prompt structure:

Positive prompt (quality tags at the start of prompt):

masterpiece, best quality, very aesthetic, {{tags}}

Slightly updated dataset of 96 images, trained at 1024 x 1024 and 1536 x 1024 resolutions, previews are mostly generated at 1536 x 1024 or 1024 x 1536 .

Used diffusion-pipe - fork by @bluvoll

Config:

# Resolution settings.
resolutions = [[1280, 720], [1536, 1024]]

# Aspect ratio bucketing settings
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 7

[[directory]] # IMAGES
# Path to the directory containing images and their corresponding caption files.
path = '/mnt/d/training_data/images'
num_repeats = 1
resolutions = [[1280, 720], [1536, 1024]]

# Config for RTX 6000 Pro 96GB
# Run with NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" NCCL_CUMEM_ENABLE="0" deepspeed --num_gpus=1 train.py --deepspeed --config config-anima.toml

# Change these paths
output_dir = '/mnt/d/anima/training_output'
dataset = 'dataset-anima.toml'

# training settings
epochs = 50
micro_batch_size_per_gpu = 24
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1.0
warmup_steps = 150
train_llm_adapter = true

# eval settings
eval_every_n_epochs = 1
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1

# misc settings
save_every_n_epochs = 1
checkpoint_every_n_epochs = 1
# checkpoint_every_n_minutes = 60
activation_checkpointing = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 1
steps_per_print = 1

[model]
type = 'anima'

transformer_path = '/mnt/c/models/diffusion_models/anima-preview2.safetensors'
vae_path = '/mnt/c/models/vae/qwen_image_vae.safetensors'
qwen_path = '../qwen0.6/Qwen3-0.6B/'
dtype = 'bfloat16'
timestep_sample_method = 'logit_normal'
sigmoid_scale = 1.0
shift = 3.0

# Caption Processing Options
cache_text_embeddings = false
# NOTE: Requires cache_text_embeddings = false to work!
# For cached embeddings, use cache_shuffle_num in your dataset config instead.
shuffle_tags = true
tag_delimiter = ', '
keep_first_n_tags = 5
shuffle_keep_first_n = 5
tag_dropout_percent = 0.3
protected_tags_file = './protected_tags.txt'

nl_shuffle_sentences = false
nl_keep_first_sentence = true

# 'tags' 'nl' 'mixed'
caption_mode = 'mixed'

debug_caption_processing = false
debug_caption_interval = 1000

llm_adapter_lr = 6e-5 # 1e-5 base

[adapter]
type = 'lora'
rank = 32
dtype = 'bfloat16'

# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
[optimizer]
type = 'adamw_optimi'
lr = 3.5e-4 # 2e-5 base
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-7 # 1e-8 base

LORA

Anima

by motimalu

Download (Beta) View on CivitAI