TRELLIS.2/data_toolkit/README.md

Dataset Preparation Toolkit

This toolkit provides a comprehensive pipeline for preparing 3D datasets, including downloading, processing, voxelizing, and latent encoding for SC-VAE and Flow Model training.

Step 1: Install Dependencies

Initialize the environment and install necessary dependencies:

. ./data_toolkit/setup.sh

Step 2: Initialize Metadata

Before processing, load the dataset metadata.

python data_toolkit/build_metadata.py <SUBSET> --root <ROOT> [--source <SOURCE>]

Arguments:

• SUBSET: Target dataset subset. Options: ObjaverseXL, ABO, HSSD, TexVerse (Training sets); SketchfabPicked, Toys4k (Test sets).
• ROOT: Root directory to save the data.
• SOURCE: Data source (Required if SUBSET is ObjaverseXL). Options: sketchfab, github.

Example: Load metadata for ObjaverseXL (sketchfab) and save to datasets/ObjaverseXL_sketchfab:

python data_toolkit/build_metadata.py ObjaverseXL --source sketchfab --root datasets/ObjaverseXL_sketchfab

Step 3: Download Data

Download the 3D assets to the local storage.

python data_toolkit/download.py <SUBSET> --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>]

Arguments:

• RANK / WORLD_SIZE: Parameters for multi-node distributed downloading.

Example: To download the ObjaverseXL subset:

Note: The example below sets a large WORLD_SIZE (160,000) for demonstration purposes, meaning only a tiny fraction of the dataset will be downloaded by this single process.

python data_toolkit/download.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab --world_size 160000

Attention: Some datasets may require an interactive Hugging Face login or manual steps. Please follow any on-screen instructions.

Update Metadata: After downloading, update the metadata registry:

python data_toolkit/build_metadata.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab

Step 4: Process Mesh and PBR Textures

Standardize 3D assets by dumping mesh and PBR textures. Note: This process utilizes the CPU.

# Dump Meshes
python data_toolkit/dump_mesh.py <SUBSET> --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>]

# Dump PBR Textures
python data_toolkit/dump_pbr.py <SUBSET> --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>]

# Get statisitics of the asset
python asset_stats.py --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>]

Example:

python data_toolkit/dump_mesh.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab
python data_toolkit/dump_pbr.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab
python asset_stats.py --root datasets/ObjaverseXL_sketchfab

Update Metadata:

python data_toolkit/build_metadata.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab

Step 5: Convert to O-Voxels

Convert the processed meshes and textures into O-Voxels format. Note: This process utilizes the CPU.

python data_toolkit/dual_grid.py <SUBSET> --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>] [--resolution <RESOLUTION>]

python data_toolkit/voxelize_pbr.py <SUBSET> --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>] [--resolution <RESOLUTION>]

Arguments:

• RESOLUTION: Target resolutions for O-Voxels, comma-separated (e.g., 256,512,1024). Default is 256.

Example: Convert ObjaverseXL to resolutions 256, 512, and 1024:

python data_toolkit/dual_grid.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab --resolution 256,512,1024
python data_toolkit/voxelize_pbr.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab --resolution 256,512,1024

At this point, the dataset is ready for SC-VAE Training

Step 6: Encode Latents

Encode sparse structures into latents to train the first-stage generator.

# 1. Encode Shape Latents
python data_toolkit/encode_shape_latent.py --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>] [--resolution <RESOLUTION>]

# 2. Encode PBR Latents
python data_toolkit/encode_pbr_latent.py --root <ROOT> [--rank <RANK> --world_size <WORLD_SIZE>] [--resolution <RESOLUTION>]

# 3. Update Metadata (Required before next step)
python data_toolkit/build_metadata.py <SUBSET> --root <ROOT>

# 4. Encode Sparse Structure (SS) Latents
python data_toolkit/encode_ss_latent.py --root <ROOT> --shape_latent_name <SHAPE_LATENT_NAME> [--rank <RANK> --world_size <WORLD_SIZE>] [--resolution <SS_RESOLUTION>] 

Arguments:

• RESOLUTION: Input O-Voxel resolution. Default is 1024.
• SS_RESOLUTION: Resolution for sparse structures. Default is 64.
• SHAPE_LATENT_NAME: The specific version name of the shape latent.

Example:

python data_toolkit/encode_shape_latent.py --root datasets/ObjaverseXL_sketchfab --resolution 512
python data_toolkit/encode_pbr_latent.py --root datasets/ObjaverseXL_sketchfab --resolution 512
python data_toolkit/encode_shape_latent.py --root datasets/ObjaverseXL_sketchfab --resolution 1024
python data_toolkit/encode_pbr_latent.py --root datasets/ObjaverseXL_sketchfab --resolution 1024

# Update metadata
python data_toolkit/build_metadata.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab

# Encode SS Latents
python data_toolkit/encode_ss_latent.py --root datasets/ObjaverseXL_sketchfab --shape_latent_name shape_enc_next_dc_f16c32_fp16_1024 --resolution 64

# Final Metadata Update
python data_toolkit/build_metadata.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab

Step 7: Render Image Conditions

Render multi-view images to train the image-conditioned generator. Note: This process may utilize the CPU.

python data_toolkit/render_cond.py <SUBSET> --root <ROOT> [--num_views <NUM_VIEWS>] [--rank <RANK> --world_size <WORLD_SIZE>]

Arguments:

• NUM_VIEWS: Number of views to render per asset. Default is 16.

Example:

python data_toolkit/render_cond.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab

Final Metadata Update:

python data_toolkit/build_metadata.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab