Aesthetic Quality Modifiers - Masterpiece

Trained on Anima Preview-3-base

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}}, {{natural language}}

Updated dataset of 386 images, all masterpiece tagged images trained in Kirazuri (Anima) model version 2 dataset.

Trained at 1024 x 1024, 1280 x 1280, and 1536 x 1024 resolutions.

Previews are mostly generated at 1536 x 1024 or 1024 x 1536 .

Training config:

diffusion-pipe commit b0aa4f1e03169f3280c8518d37570a448420f8be

# dataset-anima.toml
resolutions = [1024, 1280, 1536]

enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 9

# Totals
# 386 images
# 15504 samples/epoch

# 153 images
# 48 samples/image - 7344 samples/epoch
[[directory]]
path = '/mnt/d/training_data/0_masterpieces_kirazuri/1536x1536'
repeats = 16
resolutions = [1024, 1280, 1536]

# 44 images
# 48 samples/image - 2112 samples/epoch
[[directory]]
path = '/mnt/d/training_data/0_masterpieces_kirazuri/1280x1280'
repeats = 24
resolutions = [1024, 1280]

# 189 images
# 32 samples/image - 6048 samples/epoch
[[directory]]
path = '/mnt/d/training_data/0_masterpieces_kirazuri/1024x1024'
repeats = 32
resolutions = [1024]

# anima-lora.toml 
output_dir = '/mnt/d/anima/training_output/masterpieces-v5'

dataset = 'dataset-anima.toml'

# training settings
epochs = 5
# Per-resolution batch sizes
micro_batch_size_per_gpu = [[1024, 32], [1280, 24], [1536, 16]]
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1
warmup_steps = 100
lr_scheduler = 'cosine'

# misc settings
save_every_n_epochs = 1
#save_every_n_steps = 1000
#save_every_n_examples = 4096000
#checkpoint_every_n_epochs = 1
#checkpoint_every_n_minutes = 120
activation_checkpointing = true
#reentrant_activation_checkpointing = true

partition_method = 'parameters'
# partition_method = 'manual'
# partition_split = [10]

save_dtype = 'bfloat16'
caching_batch_size = 1
map_num_proc = 8
steps_per_print = 1
compile = true

[model]
type = 'anima'
transformer_path = '/mnt/c/workspace/models/diffusion_models/anima-preview3-base.safetensors'
vae_path = '/mnt/c/workspace/models/vae/qwen_image_vae.safetensors'
llm_path = '/mnt/c/workspace/models/text_encoders/qwen_3_06b_base.safetensors'
dtype = 'bfloat16'
#cache_text_embeddings = false
llm_adapter_lr = 1e-6
#timestep_sample_method = 'uniform'
flux_shift = true
multiscale_loss_weight = 0.5
sigmoid_scale = 1.3

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

[optimizer]
type = 'adamw_optimi'
lr = 4e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8