Mecha Girl - Mechabare | メカ娘・メカバレ

Trained on depictions of Exposed mechanical, robotic, or cybernetic body parts, systems, innards, etc of a cyborg, android, or other robotic characters.

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 and workflow/training information.

Trigger:

mechabare

Recommended Tags:

android
cyberpunk
cyborg
damaged
mecha musume
mechanical parts
mechanical arms
mechanical legs
mechanical eyes
mechanical ears
mechanical hands
mechanical horns
mechanical tail
mechanical wings
robot joints
ribs
science fiction
skeleton
spine

Training images rated for sensitivity, nsfw or explicit

Description

Trained on Anima Base 1

Updated dataset with a mix of natural language and tag captions

Partitioned dataset and trained at multi-res 512, 768, 1024, 1280, 1536

Training config:

# trained using diffusion-pipe commit b0aa4f1e03169f3280c8518d37570a448420f8be
# NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" NCCL_CUMEM_ENABLE="0" deepspeed --num_gpus=1 train.py --deepspeed --config anima-lora.toml --i_know_what_i_am_doing

output_dir = '/mnt/d/anima/training_output/anima-base-1-mechabare-v3'

dataset = 'dataset-anima-mechabare.toml'

# training settings
epochs = 2
# Per-resolution batch sizes
micro_batch_size_per_gpu = [[512, 64], [768, 32], [1024, 32], [1280, 24], [1536, 16]]
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1
warmup_steps = 50
lr_scheduler = 'cosine'

# misc settings
save_every_n_epochs = 1
activation_checkpointing = true

partition_method = 'parameters'

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-base-v1.0.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 = 4e-7
#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

resolutions = [512, 768, 1024, 1280, 1536]

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

# 348 total images

# images_mechabare\1536x1536\captions.json with 21 entries.
[[directory]]
path = '/mnt/d/training_data/images_mechabare/1536x1536'
resolutions = [512, 1024, 1280, 1536]

# images_mechabare\1280x1280\captions.json with 41 entries.
[[directory]]
path = '/mnt/d/training_data/images_mechabare/1280x1280'
resolutions = [512, 1024, 1280]

# images_mechabare\1024x1024\captions.json with 217 entries.
[[directory]]
path = '/mnt/d/training_data/images_mechabare/1024x1024'
resolutions = [512, 768, 1024]

# images_mechabare\768x768\captions.json with 69 entries.
[[directory]]
path = '/mnt/d/training_data/images_mechabare/768x768'
resolutions = [512, 768]