conf/x/vectorfusion.yaml

method: "vectorfusion"

image_size: 600 # canvas size
path_svg: ~  # if you want to load a svg file and train from it
num_stages: 1 # training stages, you can train x strokes, then freeze them and train another x strokes etc
skip_live: False # if skip_live then training from scratch
style: "iconography" # "iconography", "pixelart", "low-poly", "painting", "sketch", "ink"

# train
batch_size: 1
num_iter: 500 # num_iter per path group
# lr and optim
lr_stage_one:
  point: 1
  width: 0.1
  color: 0.01
  bg: 0.01
  optim:
    name: 'adam'
    betas: [ 0.9, 0.9 ]
    eps: 1e-6
  lr_schedule: True # use lr_scheduler
  schedule:
    name: 'linear'
    keep_ratio: 0.2
    decay_ratio: 0.4
lr_stage_two:
  point: 1
  width: 0.1
  color: 0.01
  bg: 0.01
  lr_schedule: True # use lr_scheduler
  optim:
    name: 'adam'
    betas: [ 0.9, 0.9 ]
    eps: 1e-6
  schedule:
    name: 'cosine'
    warmup_steps: 50
    warmup_start_lr: 0.02
    warmup_end_lr: 1.0
    cosine_end_lr: 0.4

# primitives
num_paths: 128 # number of strokes
path_schedule: 'repeat' # 'list'
schedule_each: 16 # [1, 3, 5, 7]
trainable_bg: False # set the background to be trainable
width: 3 # stroke width
num_segments: 4
segment_init: 'circle' # 'random'
radius: 20
coord_init: 'sparse' # 'random', 'naive', place the first control point
grid: 32 # divide the canvas into n grids
path_reinit: # reinitializing paths
  use: True
  freq: 50 # every 50 iterations
  stop_step: 800 # for SDS fine-tuning
  opacity_threshold: 0.05
  area_threshold: 64

# diffusion
model_id: "sd15" # sd14, sd15, sd21, sd21b, sdxl
ldm_speed_up: False
enable_xformers: True
gradient_checkpoint: False
cpu_offload: True
num_inference_steps: 50
guidance_scale: 7.5 # sdxl default 5.0
K: 6
lora_path: ~

# SDS
sds:
  im_size: 512
  guidance_scale: 100
  grad_scale: 1.0
  t_range: [ 0.05, 0.95 ]
  num_iter: 1000 # fine-tuning steps

# Live loss
use_distance_weighted_loss: True
xing_loss_weight: 0.01
# pixel loss
penalty_weight: 0.05