Training parameters (diffusion-pipe):

output_dir = 'output'
dataset = 'dataset.toml'
epochs = 1000
max_steps = 500
micro_batch_size_per_gpu = 4
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1.0
warmup_steps = 50
lr_scheduler = ‘constant’
compile = true
rex_cycle_steps = 1000
save_every_n_steps = 50
#checkpoint_every_n_minutes = 120
activation_checkpointing = ‘unsloth’
#activation_checkpointing = true
save_dtype = ‘bfloat16’
caching_batch_size = 1
[model]
type = ‘anima’
transformer_path = ‘anima-preview3.safetensors’
vae_path = ‘qwen_image_vae.safetensors’
qwen_path = ‘Qwen3-0.6B’
dtype = ‘bfloat16’
timestep_sample_method = ‘logit_normal’
sigmoid_scale = 1.0
llm_adapter_lr = 0
cache_text_embeddings = false
shuffle_tags = true
caption_mode = “tags”
tag_delimiter = ’, ’
debug_caption_processing = false
qwen_nf4 = false
[adapter]
type = ‘lora’
rank = 16
dtype = ‘bfloat16’
[optimizer]
type = ‘came’
lr = 2e-5
weight_decay = 0.05

Dataset config:

resolutions = [1024]

enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 7
[[directory]]
path = ‘1_eguchisaan’
num_repeats = 2