klein-4b-anime-clothes-stripper-nl

How to use:

Refer to the captions in Bakanayatsu/anime-clothes-stripper-nl dataset on how prompt it.

Example: Her clothes has been removed. Her body became fully exposed, revealing: medium breasts, pussy, and anus.

You can be very specific on what to remove.

Recommended: ImageScaleToTotalPixels at 1 megapixels and 64 resolution_steps. Apply lora to Flux2 Klein 4B (not base) at 4 steps.

Observations:

• Works as intended
• Lower resolutions other than 1024 (any aspect ratio) is bad
• May move limbs (Basically RNG, very different outputs between seeds)
• Due to the model size, outputs won't be consistently good compared Qwen-Image-Edit-2511

dataset_config.toml:

[general]
caption_extension = ".txt"
enable_bucket = true
bucket_no_upscale = false
batch_size = 16

[[datasets]]
resolution = 1024
num_repeats = 2 # 279 images * 2 = 558 * 15 epochs = 8370 steps
image_jsonl_file = "dataset.jsonl"
cache_directory = "cache"

training_config.toml:

[network_arguments]
network_dim = 256
network_alpha = 128
network_module = "networks.lora_flux_2"

[optimizer_arguments]
learning_rate = 1e-4
lr_scheduler = "cosine"
lr_scheduler_power = 0
optimizer_type = "came_pytorch.CAME"
optimizer_args = [ "weight_decay=0.01", "enable_cautious_update=True", "enable_cautious_weight_decay=True", "enable_stochastic_rounding=True", "enable_8bit=True"]

[training_arguments]
model_version = "klein-base-4b"
fp8_base = true
fp8_scaled = true
fp8_text_encoder = true
lr_warmup_steps = 0.1
max_train_epochs = 15
seed = 42
timestep_sampling = "flux2_shift"
weighting_scheme = "none"
mixed_precision = "bf16"
flash_attn = true
gradient_checkpointing = true
sdpa = true
max_data_loader_n_workers = 4
persistent_data_loader_workers = true

[sampling]
sample_every_n_epochs = 1
sample_prompts = "prompts.txt"
sample_at_first = true

[saving_arguments]
save_every_n_epochs = 1
wandb_api_key = ""
output_dir = ""
output_name = ""
wandb_run_name = ""
log_prefix = ""
log_with = "wandb"