conf/x/lowpoly.yaml

image_size: 600 # canvas size
path_svg: ~  # if you want to load a svg file and train from it
color_init: 'rand' # if skip_live=True, then use color_init to init target_img
style: "low-poly" # "iconography", "pixelart", "low-poly", "painting", "sketch", "ink"

# stable diffusion in SIVE stage
sive_model_cfg:
  model_id: "sd21b" # sd14, sd15, sd21, sd21b, sdxl
  ldm_speed_up: False
  enable_xformers: True
  gradient_checkpoint: False
  cpu_offload: True
  num_inference_steps: 100
  guidance_scale: 7.5 # sdxl default 5.0
  lora_path: ~

# lr and optim
sive_stage_optim:
  point: 1 # control points
  width: 0.1 # stroke width
  color: 0.01 # fill color and stroke color
  bg: 0.01 # bg in render_warp
  optim:
    name: 'adam'
    betas: [ 0.9, 0.9 ]
    eps: 1e-6
  schedule:
    name: 'linear'
    keep_ratio: 0.2
    decay_ratio: 0.4

# SIVE rendering
sive:
  attn_cfg: # init content via attn
    cross_attn_res: 16
    self_attn_res: 32
    max_com: 20
    mean_comp: False
    comp_idx: 0
    attn_coeff: 1.0 # attn fusion, w * cross-attn + (1-w) * self-attn
  mask_tau: 0.3 # the threshold used to convert the attention map into a mask
  bg:
    style: "iconography" # 'iconography' ,"pixelart", "sketch", 'painting', 'ink'
    num_iter: 10
    num_paths: 256
    path_schedule: 'repeat' # 'repeat', 'list'
    schedule_each: 128
    width: 3 # sketch stroke width
    num_segments: 4
    segment_init: 'circle' # 'random'
    radius: 20
    coord_init: 'random' # 'sparse', 'random', 'naive'. place the first control point
    grid: 20
    # optim
    lr_schedule: True
    optim_bg: False # train background
    use_attn_init: True
    softmax_tau: 0.3 # temperature of softmax
    # loss
    use_distance_weighted_loss: False
    xing_loss_weight: 0.001
  fg:
    style: "iconography" # 'iconography' ,"pixelart", "sketch", 'painting', 'ink'
    num_iter: 10
    num_paths: 256 # number of strokes
    path_schedule: 'repeat' # 'repeat', 'list'
    schedule_each: 128
    width: 3 # sketch stroke width
    num_segments: 4
    segment_init: 'circle' # 'random'
    radius: 15
    coord_init: 'random' # 'random', 'naive', place the first control point
    grid: 20
    # optim
    lr_schedule: False
    optim_bg: False # train background
    use_attn_init: True
    softmax_tau: 0.3 # temperature of softmax
    # loss
    use_distance_weighted_loss: False
    xing_loss_weight: 0.01
  tog: # for refinement
    reinit: False # if False, use fg params to init content
    num_iter: 10

# VPSD primitives
num_paths: 512 # number of strokes
trainable_bg: False # set the background to be trainable
width: 3 # stroke width
num_segments: 4
segment_init: 'circle' # 'random'
radius: 20
coord_init: 'random' # 'random', 'naive', 'sparse' place the first control point
grid: 30 # divide the canvas into n grids
path_reinit: # reinitializing paths
  use: True
  freq: 100 # every n iterations
  stop_step: 1000 # for VPSD fine-tuning
  opacity_threshold: 0.05
  area_threshold: 64

# lr and optim
vpsd_stage_optim:
  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: 10
    warmup_start_lr: 0.02
    warmup_end_lr: 0.9
    cosine_end_lr: 0.4

# stable diffusion in VPSD stage
vpsd_model_cfg:
  model_id: "sd21b" # sd14, sd15, sd21, sd21b, sdxl
  ldm_speed_up: False
  enable_xformers: True
  gradient_checkpoint: False
  cpu_offload: True
  num_inference_steps: 100
  guidance_scale: 7.5 # sdxl default 5.0
  lora_path: ~

# VPSD setting
vpsd:
  use: True
  type: 'vpsd'
  n_particle: 6 # 4, 8, 16
  vsd_n_particle: 4 # the batch size of particles
  particle_aug: False # do data enhancement for the input particles
  num_iter: 2000 # total iterations
  guidance_scale: 7.5 # CFG value
  grad_scale: 1.0 # increase or decrease the gradient
  grad_clip_val: ~ # eg: 10, clip the gradient of VPSD
  t_range: [ 0.02, 0.98 ]
  # 'randint': random time steps, this may have a more authentic style.
  # 'max_0.5_900': annealing from 0.98 to 0.5 after 900 steps, this may have a more colorful results.
  t_schedule: 'max_0.5_1000' # or 'randint'
  # phi model config
  phi_single: False # if False new an unet model to estimate noise
  phi_model: 'lora' # 'lora', 'unet_simple'
  use_attn_scale: ${x.vpsd.phi_single} # use lora_attn_scale or not
  lora_attn_scale: 1.0 # the scale of the attn based lora layer
  phi_guidance_scale: 1.0
  phi_t: False # different t for phi fine-tuning
  phi_update_step: 1 # enable multi-update phi model or not
  phi_lr: 0.0001 # learning rate of phi model
  phi_scheduler: 'ddim'
  phi_n_particle: 2 # the batch size of phi_model
  # ReFL config
  phi_ReFL: False # enable reward feed back learning
  n_phi_sample: 1 # number of samples used in ReFL
  phi_sample_step: 200 # the phi log step
  phi_infer_step: 50 # the phi num_inference_steps
  # phi model optim
  phi_optim:
    name: 'adamw'
    betas: [ 0.9, 0.999 ]
    eps: 1e-8
    weight_decay: ~ # 1e-5
  # phi model lr learning schedule
  phi_schedule:
    use: False
    name: 'cosine'
    warmup_steps: 50
    warmup_start_lr: 0.00001
    warmup_end_lr: 0.0001
    total_step: 800
    cosine_end_lr: 0.0001

# reward model
reward_path: './checkpoint/ImageReward'

# xing loss for closed-form paths
xing_loss:
  use: False
  weight: 0.01