GenSplat

This repository keeps three feed-forward Gaussian reconstruction variants under one training pipeline so they can be trained and compared directly.

Source layout:

• src/model/encoder/anysplat.py: preserved VGGT student + VGGT teacher
• src/model/encoder/gensplat.py: main VGGT student + Pi3 teacher
• src/model/encoder/gensplatpi3.py: comparison Pi3 student + Pi3 teacher

0. Environment Setup

This repository follows the same environment style as AnySplat, with one extra dependency required by the Pi3-based paths: utils3d.

The commands below are a practical Linux baseline for Python 3.10, PyTorch 2.2, and CUDA 12.1.

conda create -y -n gensplat python=3.10
conda activate gensplat

pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 --index-url https://download.pytorch.org/whl/cu121

pip install \
  numpy==1.25.0 wheel tqdm lightning black ruff hydra-core jaxtyping beartype \
  einops colorama scikit-image colorspacious matplotlib moviepy imageio timm \
  dacite lpips e3nn plyfile tabulate svg.py scikit-video swanlab tyro viser \
  pycolmap opencv-python Pillow huggingface_hub gradio xformers==0.0.24 \
  torch_scatter==2.1.2 pydantic open3d safetensors utils3d

pip install git+https://github.com/facebookresearch/pytorch3d.git
pip install https://github.com/nerfstudio-project/gsplat/releases/download/v1.4.0/gsplat-1.4.0%2Bpt22cu121-cp310-cp310-linux_x86_64.whl

Notes:

• If your CUDA or PyTorch version differs, adjust the torch, xformers, torch_scatter, and gsplat wheels accordingly.
• The training entry point is src/main.py.
• The default training preset in this repository is +experiment=droid.

1. Training the Three Variants

1.1 Common Preparation

Before launching training:

1. Set the dataset root in config/dataset/droid.yaml, or override it from the command line.
2. Decide whether you want to initialize from an existing checkpoint or start from the default Hugging Face weights.
3. Be explicit about checkpointing.load, because config/experiment/droid.yaml contains a sample checkpoint path.

Useful checkpoint arguments:

• model.encoder.pretrained_weights: initialize the student or the full model from a local checkpoint
• model.encoder.distiller_pretrained_weights: initialize the Pi3 teacher in the gensplat variant
• checkpointing.load: resume or override from an existing full checkpoint

If you do not want to resume from an existing run, set:

checkpointing.load=null

1.2 AnySplat: VGGT Student + VGGT Teacher

This is the preserved original baseline.

CUDA_VISIBLE_DEVICES=0,1,2,3 python src/main.py \
  +experiment=droid \
  model/encoder=anysplat \
  checkpointing.load=null \
  model.encoder.pretrained_weights=/path/to/anysplat_model.safetensors \
  swanlab.mode=offline

If you want to start from the default VGGT Hugging Face weights only, omit model.encoder.pretrained_weights.

1.3 GenSplat: VGGT Student + Pi3 Teacher

This is the main variant in this repository.

Recommended initialization:

• student initialized from an AnySplat checkpoint
• teacher initialized from a Pi3 checkpoint

CUDA_VISIBLE_DEVICES=0,1,2,3 python src/main.py \
  +experiment=droid \
  model/encoder=gensplat \
  checkpointing.load=null \
  model.encoder.pretrained_weights=/path/to/anysplat_model.safetensors \
  model.encoder.distiller_pretrained_weights=/path/to/pi3_checkpoint.pt \
  swanlab.mode=offline

If model.encoder.distiller_pretrained_weights is omitted, the teacher falls back to yyfz233/Pi3 through Pi3.from_pretrained(...).

1.4 GenSplatPi3: Pi3 Student + Pi3 Teacher

This variant is kept as a comparison model.

CUDA_VISIBLE_DEVICES=0,1,2,3 python src/main.py \
  +experiment=droid \
  model/encoder=gensplatpi3 \
  checkpointing.load=null \
  swanlab.mode=offline

You can still provide model.encoder.pretrained_weights or checkpointing.load if you have a compatible full checkpoint.

1.5 Output Directory

With +experiment=droid, training outputs are written under:

output/exp_<swanlab.name>/<timestamp>/

The default checkpoint directory is:

output/exp_<swanlab.name>/<timestamp>/checkpoints/

2. Using augment_parquet.py

The repository includes augment_parquet.py for parquet-based data augmentation with single-view GenSplat novel-view synthesis.

The script:

• reads all .parquet files from --input_dir
• processes observation.image1 and observation.image2 independently
• reconstructs each image as a single-view Gaussian scene
• perturbs the predicted camera translation by --trans_noise
• renders a new view
• writes the rendered PNG bytes back to the parquet rows
• keeps observation.image3 unchanged

2.1 Example Command

python augment_parquet.py \
  --input_dir /data/chunk-000 \
  --output_dir /data/chunk-000-new \
  --ckpt /path/to/gensplat_hf_export \
  --trans_noise 0.02

Arguments:

• --input_dir: directory containing input parquet files
• --output_dir: directory for augmented parquet files; defaults to <input_dir>-new
• --ckpt: model path or Hugging Face identifier accepted by GenSplat.from_pretrained(...)
• --trans_noise: translation noise magnitude applied to the predicted pose before rendering

2.2 Input Assumptions

The script currently assumes each parquet row contains these columns:

• observation.image1
• observation.image2
• observation.image3

It also assumes each of these columns is dict-like and contains at least:

• ["bytes"]: encoded image bytes

The script center-crops each processed image so that height and width are multiples of 14, which matches the model patch size.

3. augment_parquet.py Output Structure

The output directory mirrors the input directory at the file level:

input_dir/
  part-000.parquet
  part-001.parquet

output_dir/
  part-000.parquet
  part-001.parquet

For every output parquet file:

• the file name is unchanged
• the row count is unchanged
• all non-image columns are preserved
• observation.image3 is preserved
• observation.image1["bytes"] is replaced by rendered PNG bytes
• observation.image2["bytes"] is replaced by rendered PNG bytes

The schema is preserved; only the payload inside the selected image columns is updated.

3.1 Per-Row Behavior

Each row keeps the same top-level layout:

row
|- observation.image1
|  \- bytes   <- replaced
|- observation.image2
|  \- bytes   <- replaced
\- observation.image3
   \- bytes   <- unchanged

If observation.image1 or observation.image2 contains additional metadata keys, the script keeps that dictionary and only overwrites the bytes field.

4. Recommended Workflow

For a clean comparison:

1. Train anysplat as the preserved baseline.
2. Train gensplat with the AnySplat student checkpoint and the Pi3 teacher checkpoint.
3. Train gensplatpi3 as the Pi3-only comparison model.
4. Run augment_parquet.py only after you have a trained gensplat checkpoint exported in a format compatible with GenSplat.from_pretrained(...).

5. Relevant Files

• src/main.py: training entry point
• config/experiment/droid.yaml: main training preset
• config/model/encoder/anysplat.yaml: AnySplat encoder config
• config/model/encoder/gensplat.yaml: GenSplat encoder config
• config/model/encoder/gensplatpi3.yaml: GenSplatPi3 encoder config
• augment_parquet.py: parquet augmentation utility