LGTM

Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting

Yixing Lao^1,2†, Xuyang Bai², Xiaoyang Wu¹, Nuoyuan Yan², Zixin Luo², Tian Fang²,
Jean-Daniel Nahmias², Yanghai Tsin², Shiwei Li^2‡, Hengshuang Zhao¹

¹HKU    ²Apple
^†Work done during an internship at Apple    ^‡Project lead

ICLR 2026

Paper (PDF) | Samples Page

Existing feed-forward Gaussian Splatting methods can't scale to 4K. LGTM is the first native 4K feed-forward method that predicts compact textured Gaussians.

---

Installation

# Follow uv installation guides
# https://docs.astral.sh/uv/getting-started/installation/
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone the repository
git clone https://github.com/apple/ml-lgtm.git
cd ml-lgtm

# Install lgtm (editable) and dependencies (regular)
# This creates .venv/ automatically
uv sync
source .venv/bin/activate

# Clone, patch, and install gsplat with lgtm support
# LGTM implements custom textured 2DGS for gsplat
./scripts/clone_patch_gsplat.sh
uv pip install extern/gsplat/ --no-build-isolation

If you need a specific CUDA version, override after the uv sync step, e.g.,:

uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

---

Inference

The inference scripts download a test scene from the gated DL3DV-Benchmark HuggingFace dataset. Before running them, accept the dataset license on the linked page and log in:

hf auth login

The inference scripts will load LGTM Lightning checkpoints from weights/ and run feed-forward reconstruction on example images. Make sure you have downloaded the weights:

wget -P weights \
  https://ml-site.cdn-apple.com/models/lgtm/dl3dv_depthsplat_lgtm_512x960_2048x3840.ckpt \
  https://ml-site.cdn-apple.com/models/lgtm/dl3dv_flash3d_lgtm_288x512_2304x4096.ckpt \
  https://ml-site.cdn-apple.com/models/lgtm/dl3dv_noposplat_lgtm_288x512_2304x4096.ckpt

Now, run the inference scripts:

python examples/inference_depthsplat.py
python examples/inference_noposplat.py
python examples/inference_flash3d.py

This will render the target views and interpolation videos, along with the context images and target ground truth. You shall get something like:

examples/outputs_lgtm_depthsplat
|-- context_0.png
|-- context_9.png
|-- interpolation.mp4
|-- target_gt_1.png
|-- target_gt_3.png
|-- target_gt_5.png
|-- target_gt_7.png
|-- target_pd_1.png
|-- target_pd_3.png
|-- target_pd_5.png
`-- target_pd_7.png

---

Training

Datasets

For training, we support DL3DV dataset. We provide data processing scripts to convert the original datasets to PyTorch chunk files which can be directly loaded with this codebase. Please refer to docs/DATASETS.md for more details.

Backbone Weights

Download backbone pretrained weights (required for training):

python scripts/download_pretrained.py

You shall get:

pretrained/
|-- MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth            # For NoPoSplat
|-- depthsplat-gs-base-re10kdl3dv-448x768-randview2-6-f8ddd845.pth  # For DepthSplat
`-- unidepth-v1-vitl14.bin                                          # For Flash3D

By using backbone weights, you agree to comply with the licenses of NoPoSplat, DepthSplat, and Flash3D, along with their associated dependencies, including but not limited to: MASt3R, UniDepth, and UniMatch. See src/lgtm/model/backbone for more details of the backbone implementations.

Training Steps

NoPoSplat LGTM:

# Stage 1: 2DGS
python main.py +experiment=dl3dv_noposplat_2dgs_288x512_2304x4096
# Stage 2: BBSplat (no texture projection)
python main.py +experiment=dl3dv_noposplat_bbsplat_288x512_2304x4096
# Stage 3: LGTM
python main.py +experiment=dl3dv_noposplat_lgtm_288x512_2304x4096

DepthSplat LGTM:

# Stage 1: 2DGS
python main.py +experiment=dl3dv_depthsplat_2dgs_512x960_2048x3840
# Stage 2: LGTM
python main.py +experiment=dl3dv_depthsplat_lgtm_512x960_2048x3840

Flash3D LGTM:

# Stage 1: 2DGS
python main.py +experiment=dl3dv_flash3d_2dgs_288x512_2304x4096
# Stage 2: LGTM
python main.py +experiment=dl3dv_flash3d_lgtm_288x512_2304x4096

---

Evaluation

Prepare the test dataset in data/dl3dv/ directory:

data/dl3dv/test/
|-- 000000.torch
|-- 000001.torch
|-- ...
`-- index.json

Download the weights to weights/ directory:

wget -P weights \
  https://ml-site.cdn-apple.com/models/lgtm/dl3dv_depthsplat_lgtm_512x960_2048x3840.ckpt \
  https://ml-site.cdn-apple.com/models/lgtm/dl3dv_flash3d_lgtm_288x512_2304x4096.ckpt \
  https://ml-site.cdn-apple.com/models/lgtm/dl3dv_noposplat_lgtm_288x512_2304x4096.ckpt

Run evaluation on the test set using the weights:

# NoPoSplat LGTM
python main.py +experiment=dl3dv_noposplat_lgtm_288x512_2304x4096 mode=test wandb.mode=disabled \
    checkpointing.test_ckpt_path=weights/dl3dv_noposplat_lgtm_288x512_2304x4096.ckpt

# DepthSplat LGTM
python main.py +experiment=dl3dv_depthsplat_lgtm_512x960_2048x3840 mode=test wandb.mode=disabled \
    checkpointing.test_ckpt_path=weights/dl3dv_depthsplat_lgtm_512x960_2048x3840.ckpt

# Flash3D LGTM
python main.py +experiment=dl3dv_flash3d_lgtm_288x512_2304x4096 mode=test wandb.mode=disabled \
    checkpointing.test_ckpt_path=weights/dl3dv_flash3d_lgtm_288x512_2304x4096.ckpt

---

Citation

@inproceedings{lao2026lgtm,
  title     = {Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting},
  author    = {Lao, Yixing and Bai, Xuyang and Wu, Xiaoyang and Yan, Nuoyuan and Luo, Zixin and Fang, Tian and Nahmias, Jean-Daniel and Tsin, Yanghai and Li, Shiwei and Zhao, Hengshuang},
  booktitle = {ICLR},
  year      = {2026},
}

License

Please check out the repository LICENSE before using the provided code and LICENSE_MODEL for the released models.