Trained on Anima Preview

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

Updated dataset and reduced to 78 images, adjusted captions for Anima with a mix of NL and tags.

Used diffusion-pipe - fork by @bluvoll

Config:

# dataset-anima.toml
# Resolution settings.
resolutions = [1024]

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

# config-anima.toml
[[directory]] # IMAGES
# Path to the directory containing images and their corresponding caption files.
path = '/mnt/d/huanvideo/training_data/images'
num_repeats = 1
resolutions = [1024]

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

# training settings
epochs = 50
micro_batch_size_per_gpu = 6
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1.0
warmup_steps = 100

# 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_minutes = 120
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-preview.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 = 0.10
protected_tags_file = './protected_tags.txt'

nl_shuffle_sentences = true
nl_keep_first_sentence = false

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

debug_caption_processing = true
debug_caption_interval = 100

[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 = 5e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8