Sparse tensor with support for both torchsparse and spconv backends. Parameters: - feats (torch.Tensor): Features of the sparse tens

Renderer for the Mesh representation. Args: rendering_options (dict): Rendering options.

Sequential tensor with variable length. Args: feats (torch.Tensor): Features of the varlen tensor. layout (List[slice]):

Distributed sampler that is resumable and balances the load among the processes. Args: dataset: Dataset used for sampling. r

DINOv3 Feature Extractor with View-Aligned Projection. This extractor produces both: 1. Global features (CLS token + register tokens

Renderer for the Voxel representation. Args: rendering_options (dict): Rendering options.

Embeds spatial positions into vector representations.

Renderer for the PBR mesh. Args: rendering_options (dict): Rendering options.

3D Grid Projection Module. Projects a 3D grid of points to 2D image coordinates and samples features from the image feature map at

Embeds scalar timesteps into vector representations.

DINOv3 + Flux VAE Feature Extractor with Gated Fusion and View-Aligned Projection. Produces three outputs for GatedProjectAttention:

Concat-Projection Attention Module for DINOv3 (semantic) + VAE (color) features. Concatenates DINOv3 and VAE projected features and appl

A GroupNorm layer that converts to float32 before the forward pass.

Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning. Supports two image attention modes:

Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning. Supports two image attention modes: - "cr

Projection-based Attention Module with per-block proj_linear. Combines global cross-attention with view-aligned projected features.

Distributed sampler that is resumable. Args: dataset: Dataset used for sampling. rank (int, optional): Rank of the current p

Sparse Concat-Projection Attention Module for DINOv3 + VAE features. Concatenates DINOv3 and VAE projected features and applies a single

Sparse Projection-based Attention Module with per-block proj_linear.

Adaptive gradient clipping for training.

Trainer for basic training loop. Args: models (dict[str, nn.Module]): Models to train. dataset (torch.utils.data.Dataset

A mixin class for samplers that apply classifier-free guidance.

Feature extractor for DINOv2 models.

Feature extractor for DINOv2 models.

Feature extractor for DINOv3 models.

Feature extractor for DINOv3 models.

Module for training with elastic memory management.

Mixin for training with elastic memory management.

SLat Flow Model with elastic memory management. Used for training with low VRAM.

Flexible Dual Grid Dataset Args: roots (str): path to the dataset resolution (int): resolution of the voxel grid

Generate samples from a flow-matching model using Euler sampling with classifier-free guidance.

Generate samples from a flow-matching model using Euler sampling with classifier-free guidance and interval.

Generate samples from a flow-matching model using Euler sampling. Args: sigma_min: The minimum scale of noise in flow.

Trainer for diffusion model with flow matching objective and classifier-free guidance. Args: models (dict[str, nn.Module]): Mode

Trainer for diffusion model with flow matching objective. Args: models (dict[str, nn.Module]): Models to train. dataset

A mixin class for samplers that apply classifier-free guidance with interval.

Trainer for image-conditioned diffusion model with flow matching objective and classifier-free guidance. Args: models (dict[str,

Mixin for image-conditioned models. Args: image_cond_model: The image conditioning model.

Trainer for image-conditioned diffusion model with view-aligned projection. Uses ImageConditionedProjMixin for 3D-to-2D feature projecti

Mixin for image-conditioned models with view-aligned projection. This mixin adds support for extracting view-aligned features from image

Trainer for sparse image-conditioned diffusion model with view-aligned projection. Uses ImageConditionedProjMixin for 3D-to-2D feature p

Image conditioned structured latent dataset

Image conditioned structured latent dataset

Image conditioned structured latent for shape generation

Trainer for sparse image-conditioned diffusion model with flow matching objective and classifier-free guidance. Args: models (di

Image-conditioned sparse structure dataset

A simple controller for memory management during training. The memory usage is modeled as a linear function of: - the number of input

Base class for memory management during training.

Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.

Transformer block (MSA + FFN) with adaptive layer norm conditioning.

Mixin for multiple-image-conditioned models. Args: image_cond_model: The image conditioning model.

Trainer for sparse image-conditioned diffusion model with flow matching objective and classifier-free guidance. Args: models (di

Trainer for PBR attributes VAE Args: models (dict[str, nn.Module]): Models to train. dataset (torch.utils.data.Dataset):

A base class for pipelines.

Pipeline for inferring Pixal3D (proj mode) image-to-3D models. Based on Trellis2 pipeline, using proj mode for inference. Each stage (SS

structured latent V2 dataset Args: roots (str): path to the dataset min_aesthetic_score (float): minimum aesthetic score

Structured Latent Flow Model for 3D generation. Supports two conditioning modes: - "cross": Standard cross-attention with image feat

structured latent for sparse voxel pbr dataset Args: roots (str): path to the dataset latent_key (str): key of the laten

View-based structured latent for PBR/texture generation with view-aligned projection. Data format: PBR latent: {sha256}/view{

structured latent for shape generation Args: roots (str): path to the dataset resolution (int): resolution of the shape

View-based structured latent for shape generation. Data format: {sha256}/view{XX}.npz where each npz contains 'coords' and 'feats' keys.

A base class for samplers.

Trainer for Shape VAE Args: models (dict[str, nn.Module]): Models to train. dataset (torch.utils.data.Dataset): Dataset.

Upsample a sparse tensor by a factor of `factor`. Implemented as rearranging its features from channel to spatial.