refiners/configs/finetune-textual-inversion.toml

[wandb]
mode = "offline" # "online", "offline", "disabled"
entity = "acme"
project = "test-textual-inversion"

[models]
unet = {checkpoint = "/path/to/stable-diffusion-1-5/unet.safetensors"}
text_encoder = {checkpoint = "/path/to/stable-diffusion-1-5/CLIPTextEncoderL.safetensors"}
lda = {checkpoint = "/path/to/stable-diffusion-1-5/lda.safetensors"}

[latent_diffusion]
unconditional_sampling_probability = 0.05
offset_noise = 0.1

[textual_inversion]
placeholder_token = "<cat-toy>"
initializer_token = "toy"
# style_mode = true

[training]
duration = "2000:step"
seed = 0
gpu_index = 0
batch_size = 4
gradient_accumulation = "1:step"
evaluation_interval = "250:step"
evaluation_seed = 1

[optimizer]
optimizer = "AdamW" # "SGD", "Adam", "AdamW", "AdamW8bit", "Lion8bit"
learning_rate = 5e-4
betas = [0.9, 0.999]
eps = 1e-8
weight_decay = 1e-2

[scheduler]
scheduler_type = "ConstantLR"
update_interval = "1:step"

[dropout]
dropout_probability = 0
use_gyro_dropout = false

[dataset]
hf_repo = "acme/images"
revision = "main"
horizontal_flip = true
random_crop = true
resize_image_max_size = 512

[checkpointing]
# save_folder = "/path/to/ckpts"
save_interval = "250:step"

[test_diffusion]
num_inference_steps = 30
use_short_prompts = false
prompts = [
    "<cat-toy>",
    # "green grass, <cat-toy>"
]