Initialize models and more.

Resets the scene to a clean state. Returns: None

Resets the scene to a clean state. Returns: None

Resets the scene to a clean state. Returns: None

OpenCV intrinsics to OpenGL perspective matrix Args: intrinsics (torch.Tensor): [3, 3] OpenCV intrinsics matrix near (float)

OpenCV intrinsics to OpenGL perspective matrix Args: intrinsics (torch.Tensor): [3, 3] OpenCV intrinsics matrix near (float)

Standard linear-to-sRGB conversion matching Blender's sRGB display transform. Applies the official sRGB EOTF inverse: - For values

Load data.

Load HDRI environment maps from assets/.

Loads a model with a supported file extension into the scene. Args: object_path (str): Path to the model file. Raises: Value

Loads a model with a supported file extension into the scene. Args: object_path (str): Path to the model file. Raises: Value

Loads a model with a supported file extension into the scene. Args: object_path (str): Path to the model file. Raises: Value

Copy model parameters into a inflated tensor of full-precision parameters.

Copy the model parameter data into the master parameters.

Normalizes the scene by scaling and translating it to fit in a unit cube centered at the origin. Mostly taken from the Point-E / Shap-E rende

Normalizes the scene by scaling and translating it to fit in a unit cube centered at the origin. Mostly taken from the Point-E / Shap-E rende

Normalizes the scene by scaling and translating it to fit in a unit cube centered at the origin. Mostly taken from the Point-E / Shap-E rende

Apply PCA to the features and return the first three principal components.

3D pixel shuffle.

Prepare dataloader.

Prepare PBR dump data for processing. Consistent with voxelize_pbr.py preprocessing. Args: dump: raw PBR dump data Returns:

Preprocess the input image. Args: input: Input image (RGB or RGBA). bg_color: Background color (R, G, B) in

Preprocess the input image.

Preprocess the input mesh.

Profile the training loop.

Convert proj camera parameters to renderer-compatible extrinsics + intrinsics. The proj camera (Blender convention) views from the front

Render a turntable video of the sample. Args: start_yaw: Starting yaw angle in radians. If None, defaults to π/2. start_

Sample features from feature map at specified NDC coordinates. Args: fmap: torch.Tensor, shape [B, C, H, W], feature map

Sample x_{t-1} from the model using Euler method. Args: model: The model to sample from. x_t: The [N

Sample structured latent with the given conditioning. Args: cond (dict): The conditioning information.

Sample sparse structures with the given conditioning. Args: cond (dict): The conditioning information.

Sample sparse structures with the given conditioning. Args: cond (dict): The conditioning information.

Sample texture structured latent with the given conditioning. Args: cond (dict): The conditioning information.

Sample structured latent with the given conditioning. Args: cond (dict): The conditioning information.

Save an image with notes.

Screen space ambient occlusion (SSAO) Args: depth (torch.Tensor): [H, W, 1] depth image normal (torch.Tensor): [H, W, 3] nor

Sample images from the dataset.

Concatenate a list of sparse tensors. Args: inputs (List[SparseTensor]): List of sparse tensors to concatenate.

Unbind a sparse tensor along a dimension. Args: input (SparseTensor): Sparse tensor to unbind. dim (int): Dimension to u

Create sinusoidal timestep embeddings. Args: t: a 1-D Tensor of N indices, one per batch element. These

Compute training losses.

Apply multi-view transform to mesh vertices based on camera transform matrix.

Apply multi-view transform to normals (rotation only). Consistent with test_ovoxel_pbr_transform.py implementation. Args: normal

Apply multi-view transform to entire PBR dump data. Processing flow (based on test_ovoxel_pbr_transform.py): 1. Box normalize all vertic

Apply multi-view transform to vertices based on camera transform matrix. Args: vertices: torch.Tensor, shape [N, 3], vertex coordina

Tries to extract a factor from a math node link. Supported sub tree modes: - color: ANY -> MIX(MULTIPLY) -> - scalar:

Tries to extract an image from a texture node link. Supported sub tree modes: - RGB: TEX_IMAGE -> - R, G, B: TEX_

Unflatten the master parameters to look like model_params.

Update exponential moving average. Should only be called by the rank 0 process.

Unbind a varlen tensor along a dimension. Args: input (VarLenTensor): Varlen tensor to unbind. dim (int): Dimension to u

Context manager for training with a reduced memory ratio compared to the full memory usage. Returns: float: The

Wrap a single line of text into multiple lines such that each line's width (measured at the given scale) does not exceed max_width.