Argus Forge

# argus-forge (demo) dataset config for kohya sd-scripts (pass via --dataset_config)
# 40 images x 4 repeats x 10 epochs = 1600 samples (800 optimizer steps @ batch 2)

[general]
enable_bucket = true
caption_extension = ".txt"
shuffle_caption = false
keep_tokens = 0

[[datasets]]
resolution = 1024
batch_size = 2
min_bucket_reso = 256
max_bucket_reso = 2048
bucket_reso_steps = 64

[[datasets.subsets]]
# Demo placeholder — point this at your exported dataset.
image_dir = "/data/out"
num_repeats = 4
# Fallback caption for images without a .txt sidecar.
class_tokens = "subject"

# argus-forge (demo) training config for kohya sd-scripts (pass via --config_file)
# Seeded from argus-curator selection insights (identity, 40 images).
# Starting points, not gospel — watch samples and stop early if it overfits.

pretrained_model_name_or_path = "stabilityai/stable-diffusion-xl-base-1.0"
output_dir = "/data/out/forge/kohya/output"
output_name = "subject-lora"
save_model_as = "safetensors"
save_every_n_epochs = 1
save_precision = "bf16"

max_train_epochs = 10
train_batch_size = 2
seed = 42
mixed_precision = "bf16"

network_module = "networks.lora"
network_dim = 16
network_alpha = 8

learning_rate = 1e-4
unet_lr = 1e-4
text_encoder_lr = 5e-5
optimizer_type = "AdamW8bit"
lr_scheduler = "cosine"
lr_warmup_steps = 40
min_snr_gamma = 5

gradient_checkpointing = true
cache_latents = true
sdpa = true
# SDXL's fp16 VAE is numerically unstable; keep it in fp32.
no_half_vae = true
logging_dir = "/data/out/forge/kohya/logs"