debug: False
root_data_dir: ./runs
checkpoint_path: 
global_seed: 42

inference:
  input_dir: 
  ddim: 25
  cfg: 6
  seed: ${global_seed}
  precision: fp16
  export_glb: True
  decimate: 100000
  mc_resolution: 256
  batch_size: 4096
  remesh: False

image_height: 518
image_width: 518

model:
  class_name: models.primsdf.PrimSDF
  num_prims: 2048
  dim_feat: 6
  prim_shape: 8
  init_scale: 0.05 # useless if auto_scale_init == True
  sdf2alpha_var: 0.005
  auto_scale_init: True
  init_sampling: uniform
  vae:
    class_name: models.vae3d_dib.VAE
    in_channels: ${model.dim_feat}
    latent_channels: 1
    out_channels: ${model.vae.in_channels}
    down_channels: [32, 256]
    mid_attention: True
    up_channels: [256, 32]
    layers_per_block: 2
    gradient_checkpointing: False
  vae_checkpoint_path: 
  conditioner:
    class_name: models.conditioner.image.ImageConditioner
    num_prims: ${model.num_prims}
    dim_feat: ${model.dim_feat}
    prim_shape: ${model.prim_shape}
    sample_view: False
    encoder_config:
      class_name: models.conditioner.image_dinov2.Dinov2Wrapper
      model_name: dinov2_vitb14_reg
      freeze: True
  generator:
    class_name: models.dit_crossattn.DiT
    seq_length: ${model.num_prims}
    in_channels: 68 # equals to model.vae.latent_channels * latent_dim^3
    condition_channels: 768
    hidden_size: 1152
    depth: 28
    num_heads: 16
    attn_proj_bias: True
    cond_drop_prob: 0.1
    gradient_checkpointing: False
  latent_nf: 1.0
  latent_mean: [ 0.0442, -0.0029, -0.0425, -0.0043, -0.4086, -0.2906, -0.7002, -0.0852, -0.4446, -0.6896, -0.7344, -0.3524, -0.5488, -0.4313, -1.1715, -0.0875, -0.6131, -0.3924, -0.7335, -0.3749,  0.4658, -0.0236,  0.8362,  0.3388,  0.0188,  0.5988, -0.1853,  1.1579,  0.6240,  0.0758,  0.9641,  0.6586,  0.6260,  0.2384,  0.7798,  0.8297, -0.6543, -0.4441, -1.3887, -0.0393, -0.9008, -0.8616, -1.7434, -0.1328, -0.8119, -0.8225, -1.8533, -0.0444, -1.0510, -0.5158, -1.1907, -0.5265,  0.2832,  0.6037,  0.5981,  0.5461,  0.4366,  0.4144,  0.7219,  0.5722,  0.5937,  0.5598,  0.9414,  0.7419,  0.2102,  0.3388,  0.4501,  0.5166]
  latent_std: [0.0219, 0.3707, 0.3911, 0.3610, 0.7549, 0.7909, 0.9691, 0.9193, 0.8218, 0.9389, 1.1785, 1.0254, 0.6376, 0.6568, 0.7892, 0.8468, 0.8775, 0.7920, 0.9037, 0.9329, 0.9196, 1.1123, 1.3041, 1.0955, 1.2727, 1.6565, 1.8502, 1.7006, 0.8973, 1.0408, 1.2034, 1.2703, 1.0373, 1.0486, 1.0716, 0.9746, 0.7088, 0.8685, 1.0030, 0.9504, 1.0410, 1.3033, 1.5368, 1.4386, 0.6142, 0.6887, 0.9085, 0.9903, 1.0190, 0.9302, 1.0121, 0.9964, 1.1474, 1.2729, 1.4627, 1.1404, 1.3713, 1.6692, 1.8424, 1.5047, 1.1356, 1.2369, 1.3554, 1.1848, 1.1319, 1.0822, 1.1972, 0.9916]

diffusion:
  timestep_respacing:
  noise_schedule: squaredcos_cap_v2
  diffusion_steps: 1000
  parameterization: v

rm:
  volradius: 10000.0
  dt: 1.0

optimizer:
  class_name: torch.optim.AdamW
  lr: 0.0001
  weight_decay: 0

scheduler:
  class_name: dva.scheduler.CosineWarmupScheduler
  warmup_iters: 3000
  max_iters: 200000

train:
  batch_size: 8
  n_workers: 4
  n_epochs: 1000
  log_every_n_steps: 50
  summary_every_n_steps: 10000
  ckpt_every_n_steps: 10000
  amp: False
  precision: tf32

tag: 3dtopia-xl-sview
output_dir: ${root_data_dir}/inference/${tag}