config.yaml

name: craftsman/craftsman-doravae
description: 'image to 3d shape diffusion of CraftsMan3D(https://github.com/wyysf-98/CraftsMan3D) and DoraVAE version(https://aruichen.github.io/Dora/)'

system_type: shape-rectified-flow-system
system:
  z_scale_factor: 1.0
  guidance_scale: 7.5
  num_inference_steps: 50
  eta: 0.0
  extract_mesh_func: mc
  shape_model_type: dora-autoencoder
  shape_model:
    n_samples: 16384
    with_sharp_data: true
    use_downsample: true
    num_latents: 2048
    embed_dim: 64
    point_feats: 3
    out_dim: 1
    num_freqs: 8
    include_pi: false
    heads: 12
    width: 768
    num_encoder_layers: 8
    num_decoder_layers: 16
    use_ln_post: true
    init_scale: 0.25
    qkv_bias: false
    use_flash: true
    use_checkpoint: true
  condition_model_type: "dinov2-embedder"
  condition_model:
    dino_type: "dinov2-with-registers-large"
    encode_camera: false
    n_views: 1
    empty_embeds_ratio: 0.1
    normalize_embeds: false
    zero_uncond_embeds: true
    image_size_dino: 518
  denoiser_model_type: pixart-denoiser
  denoiser_model:
    input_channels: 64
    output_channels: 64
    n_ctx: 2048
    width: 1280
    layers: 32
    heads: 20
    context_dim: 1024
    init_scale: 1.0
    use_checkpoint: true
    condition_type: dinov2
    use_RMSNorm: true
    qkv_fuse: false
  noise_scheduler_type: diffusers.schedulers.FlowMatchEulerDiscreteScheduler
  noise_scheduler:
    num_train_timesteps: 1000
    shift: 3.0
  denoise_scheduler_type: diffusers.schedulers.FlowMatchEulerDiscreteScheduler
  denoise_scheduler:
    num_train_timesteps: 1000
    shift: 3.0
  val_samples_json: ""
  loggers:
    wandb:
      enable: false
      project: CraftsMan
      name: image-to-shape-diffusion
  loss:
    loss_type: mse
    lambda_diffusion: 1.0
  optimizer:
    name: AdamW
    args:
      lr: 0.0002
      betas:
      - 0.9
      - 0.99
      eps: 1.0e-06
  scheduler:
    name: CosineAnnealingLR
    args:
      T_max: 5
      eta_min: 1.0e-06