# Training config for Anima Preview 2 (RECOMENDADO)
# Project: loven_base_init
# Generated: 2026-04-03 04:45:11
# Edit this file before launching training!

# === Model ===
pretrained_model_name_or_path = "/workspace/models/anima-preview2/split_files/diffusion_models/anima-preview2.safetensors"
qwen3 = "/workspace/models/anima-preview2/split_files/text_encoders/qwen_3_06b_base.safetensors"
vae = "/workspace/models/anima-preview2/split_files/vae/qwen_image_vae.safetensors"
vae_chunk_size = 64
vae_disable_cache = true
# llm_adapter_lr = 0  # uncomment for full fine-tune only
dataset_config = "/workspace/projects/loven_base_init/dataset.toml"

# === Network (LoRA) ===
network_module = "networks.lora_anima"
network_dim = 32
network_alpha = 16
network_train_unet_only = true

# === Attention ===
attn_mode = "torch"

# === Precision & Memory ===
mixed_precision = "bf16"
# full_bf16 = true  # uncomment if PyTorch >= 2.5 (saves VRAM, risk of NaN on older versions)
gradient_checkpointing = true

# === Cache ===
cache_latents = true
cache_text_encoder_outputs = true

# === Timestep ===
timestep_sampling = "sigmoid"
weighting_scheme = "uniform"
discrete_flow_shift = 3.0
sigmoid_scale = 1.0
# DO NOT set noise_offset — causes green tint artifacts on ANIMA
# To use shift-based timestep: change timestep_sampling to 'shift'

# === Optimizer ===
optimizer_type = "AdamW8bit"
learning_rate = 2e-05
lr_scheduler = "cosine"
gradient_accumulation_steps = 1
max_grad_norm = 1.0
lr_warmup_steps = 0.1

# === Duration ===
max_train_steps = 3000

# === Save ===
save_every_n_steps = 250
output_dir = "/workspace/projects/loven_base_init/outputs/20260403_044418"
output_name = "loven_base_init"
save_model_as = "safetensors"
save_precision = "bf16"

# === Sampling ===
sample_prompts = "/workspace/projects/loven_base_init/sample_prompts.txt"
sample_every_n_steps = 100

# === Dataloader ===
max_data_loader_n_workers = 2
persistent_data_loader_workers = true
seed = 42