Inconsistent Normalization Settings Between Generation and Data Processing

[xfanhe]

Dear authors,

There's a normalization mismatch in the codebase:

• Data preprocessing normalizes meshes to [0, 1] space
  

  Hunyuan3D-2.1/hy3dshape/tools/watertight/watertight_and_sample.py

  Lines 160 to 170 in 7cc51b6

  	def normalize_to_unit_box(V):
  	"""
  	Normalize the vertices V to fit inside a unit bounding box [0,1]^3.
  	V: (n,3) numpy array of vertex positions.
  	Returns: normalized V
  	"""
  	V_min = V.min(axis=0)
  	V_max = V.max(axis=0)
  	scale = (V_max - V_min).max() * 1.01
  	V_normalized = (V - V_min) / scale
  	return V_normalized

• Latent generation queries points in [-1, 1] space
  

  Hunyuan3D-2.1/hy3dshape/hy3dshape/models/autoencoders/volume_decoders.py

  Lines 157 to 167 in 7cc51b6

  	# 1. generate query points
  	if isinstance(bounds, float):
  	bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
  	bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6])
  	xyz_samples, grid_size, length = generate_dense_grid_points(
  	bbox_min=bbox_min,
  	bbox_max=bbox_max,
  	octree_resolution=octree_resolution,
  	indexing="ij"
  	)

  
  and surface_loader normalizes meshes into [-1, 1]
  

  Hunyuan3D-2.1/hy3dshape/hy3dshape/surface_loaders.py

  Lines 23 to 44 in 7cc51b6

  	def normalize_mesh(mesh, scale=0.9999):
  	"""
  	Normalize the mesh to fit inside a centered cube with a specified scale.
  	The mesh is translated so that its bounding box center is at the origin,
  	then uniformly scaled so that the longest side of the bounding box fits within [-scale, scale].
  	Args:
  	mesh (trimesh.Trimesh): Input mesh to normalize.
  	scale (float, optional): Scaling factor to slightly shrink the mesh inside the unit cube. Default is 0.9999.
  	Returns:
  	trimesh.Trimesh: The normalized mesh with applied translation and scaling.
  	"""
  	bbox = mesh.bounds
  	center = (bbox[1] + bbox[0]) / 2
  	scale_ = (bbox[1] - bbox[0]).max()
  	mesh.apply_translation(-center)
  	mesh.apply_scale(1 / scale_ * 2 * scale)
  	return mesh

Which normalization settings were used during training? Should we use [-1, 1] for both vae and diffusion training to maintain consistency?

Thank you for your excellent work!