Suzushiro Mizuki from the manga Boku Girl. Trained for Anima Preview 3.
Tags: 1girl, suzushiro mizuki, pink hair, short hair, grey eyes, mole under mouth
Optional: small breasts, flat chest
(Mole under mouth is also optional, will not show up most of the time if not tagged for).
Trained at 1024x resolution for 500 steps on diffusion-pipe (See about version for parameters).
Getting a good dataset for this character is pretty much impossible, so there may be issues with eye consistency/poses. As such, using an artist tag is recommended. Previews were generated with artist tags and upscaled.
Description
Training parameters:
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
save_every_n_steps = 50
activation_checkpointing = 'unsloth'
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
tag_dropout_percent = 0.1
caption_dropout_percent = 0.1
caption_mode = "tags"
nl_shuffle_sentences = false
tag_delimiter = ', '
protected_tags_file = 'protected_tags.txt'
debug_caption_processing = false
qwen_nf4 = false
[adapter]
type = 'lora'
rank = 16
dtype = 'bfloat16'
[optimizer]
type = 'came'
lr = 2e-5
weight_decay = 0.05Dataset config:
resolutions = [1024]
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 9
[[directory]]
path = '1_suzushiro'
num_repeats = 4