worldfm/default.yaml at main · wangnancpp/worldfm · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
# WorldFM Pipeline Default Configuration
# Override via CLI args: python run_pipeline.py --meta ... --worldfm.model_path ...
# Or load a custom config: python run_pipeline.py --config my_config.yaml --meta ...

# ─── Pipeline ─────────────────────────────────────────────────────────────────
pipeline:
  output_dir: outputs
  gpu_index: 0

# ─── External repos (auto-detect from submodules/ if empty) ──────────────────
submodules:
  hw_path: ""
  moge_path: ""

# ─── Step 1: Panorama generation (HunyuanWorld) ─────────────────────────────
panogen:
  seed: 42
  fp8_attention: false
  fp8_gemm: false
  cache: false

# ─── Step 2: MoGe depth estimation ───────────────────────────────────────────
moge:
  pretrained: Ruicheng/moge-2-vitl-normal
  resolution_level: 30
  fov_deg: 45.0
  num_views: 42
  merge_max_width: 4096
  merge_max_height: 2048
  batch_size: 4

# ─── Step 3: Condition rendering & view selection ────────────────────────────
render:
  render_size: 512
  sample_grid: 10
  center_grid: 15
  center_frac: 0.5
  eps_rel: 0.02
  eps_abs: 0.0
  px_radius: 0
  max_view_angle_deg: 180.0
  use_distance_weight: true
  dist_min_m: 1.0
  dist_max_m: 20.0
  weight_near: 1.0
  weight_far: 0.0

# ─── Step 4: WorldFM inference ───────────────────────────────────────────────
worldfm:
  step: 2
  model_path: weights/worldfm_${worldfm.step}-step.pth
  vae_path: weights/vae
  image_size: 512
  version: sigma
  cfg_scale: 4.5
  compile_model: true
  compile_mode: reduce-overhead
  vae_channels_last: true
  vae_deterministic: true
  compile_vae: true
  compile_vae_mode: reduce-overhead
  disable_vae_slicing: true
  disable_vae_tiling: true