Twunk Bastian [Flux]

apply_t5_attn_mask = true

bucket_reso_steps = 64

clip_skip = 1

discrete_flow_shift = 3.0

enable_bucket = true

epoch = 30

fp8_base = true

gradient_accumulation_steps = 1

gradient_checkpointing = true

guidance_scale = 1.0

lr_scheduler = "cosine"

lr_warmup_steps = 320

max_bucket_reso = 2048

max_grad_norm = 1

max_timestep = 1000

max_train_steps = 4800

min_bucket_reso = 256

min_snr_gamma = 7

mixed_precision = "bf16"

network_alpha = 16

network_dim = 16

network_train_unet_only = true

noise_offset = 0.05

optimizer_type = "AdamW8bit"

pretrained_model_name_or_path = "flux-dev2pro-fp8_e4m3fn_comfy.safetensors"

resolution = "768,768"

sample_sampler = "euler"

save_every_n_epochs = 1

save_model_as = "safetensors"

save_precision = "bf16"

sdpa = true

unet_lr = 0.0004