seed: 1
image_size: 224
mask_object: False # if the target image contains background, it's better to mask it out
fix_scale: False  # if the target image is not squared, it is recommended to fix the scale

# train
num_iter: 1000
batch_size: 1
num_stages: 1 # training stages, you can train x strokes, then freeze them and train another x strokes etc
lr_scheduler: False
lr_decay_rate: 0.1
decay_steps: [ 1000, 1500 ]
lr: 1
color_lr: 0.01
pruning_freq: 50
color_vars_threshold: 0.1
width_lr: 0.1
max_width: 50 # stroke width

# stroke attrs
num_paths: 96 # number of strokes
width: 1.0 # stroke width
control_points_per_seg: 4
num_segments: 1
optim_opacity: True # if True, the stroke opacity is optimized
optim_width: False  # if True, the stroke width is optimized
optim_rgba: False # if True, the stroke RGBA is optimized
opacity_delta: 0 # stroke pruning

# init strokes
attention_init: True # if True, use the attention heads of Dino model to set the location of the initial strokes
xdog_intersec: True # initialize along the edge, mix XDoG and attn up
softmax_temp: 0.5
cross_attn_res: 16
self_attn_res: 32
max_com: 20 # select the number of the self-attn maps
mean_comp: False # the average of the self-attn maps
comp_idx: 0 # if mean_comp==False, indicates the index of the self-attn map
attn_coeff: 1.0 # attn fusion, w * cross-attn + (1-w) * self-attn
log_cross_attn: False # True if cross attn every step
u2net_path: "./checkpoint/u2net/u2net.pth"

# ldm
model_id: "sd14"
ldm_speed_up: False
enable_xformers: False
gradient_checkpoint: False
#token_ind: 1 # the index of CLIP prompt embedding, start from 1
use_ddim: True
num_inference_steps: 50
guidance_scale: 7.5 # sdxl default 5.0
# ASDS loss
sds:
  crop_size: 512
  augmentations: "affine"
  guidance_scale: 100
  grad_scale: 1e-5
  t_range: [ 0.05, 0.95 ]
  warmup: 0

clip:
  model_name: "RN101"  # RN101, ViT-L/14
  feats_loss_type: "l2" # clip visual loss type, conv layers
  feats_loss_weights: [ 0,0,1.0,1.0,0 ] # RN based
  #  feats_loss_weights: [ 0,0,1.0,1.0,0,0,0,0,0,0,0,0 ] # ViT based
  fc_loss_weight: 0.1 # clip visual loss, fc layer weight
  augmentations: "affine" # augmentation before clip visual computation
  num_aug: 4 # num of augmentation before clip visual computation
  vis_loss: 1 # 1 or 0 for use or disable clip visual loss
  text_visual_coeff: 0 # cosine similarity between text and img

perceptual:
  name: "lpips" # dists
  lpips_net: 'vgg'
  coeff: 0.2