AdaptiveSplat:Texture Aware Controllable 3D Gaussian Allocation for Feed-Forward Reconstruction

Current feed-forward 3D reconstruction methods predict pixel aligned Gaussian primitives, resulting in highly redundant representations. A natural solution is to prune the redundant Gaussians, but naive pruning introduces severe artifacts and often requires inference time fine-tuning, breaking the feed-forward paradigm. Based on previous works, high frequency regions require more Gaussian primitives, while low frequency regions can be represented with significantly fewer primitives. Motivated by this, we propose a novel approach to explicitly control the number of Gaussians by leveraging local texture information. Our approach achieves this through three key components: (1) texture estimation to capture spatial variation in scene detail, (2) texture-aware pruning that removes redundant Gaussians from low frequency regions, and (3) an adaptive Gaussian head that predicts the modified attributes of the retained primitives without breaking the feed-forward paradigm. Experiments on RE10K, ACID, DL3DV, Tanks and Temples, and DTU demonstrate the effectiveness of our approach, while ablation studies validate the contributions of its key components.

Requires Linux, Python 3.10, and a CUDA-capable GPU (tested with CUDA 12.1).

# 1) Create the environment
conda create -y -n adaptive-splat python=3.10
conda activate adaptive-splat

# 2) Install PyTorch matching your CUDA toolkit
pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 \
    --index-url https://download.pytorch.org/whl/cu121

# 3) Install the remaining dependencies
pip install -r requirements.txt

# 4) Build the CroCo RoPE CUDA kernels
cd src/model/encoder/backbone/croco/curope && pip install -e . && cd -

Alternatively, use the conda spec: conda env create -f environment.yml.

Set the dataset location in the corresponding config (e.g. config/dataset/dl3dv.yaml):

roots: [datasets/dl3dv]   # path to your dataset

# Sample fully-working training command (single GPU, W&B disabled):
CUDA_VISIBLE_DEVICES=0 python -m src.main +experiment=dl3dv wandb.mode=disabled

To resume/initialize from a checkpoint, add checkpointing.load=/path/to/epoch_X-step_Y.ckpt.

# Sample test command (novel-view synthesis on DL3DV).
# Point checkpointing.load at a checkpoint produced by training above:
CUDA_VISIBLE_DEVICES=0 python -m src.main +experiment=dl3dv mode=test \
  wandb.mode=disabled \
  dataset.dl3dv.view_sampler.num_context_views=9 \
  test.global_prune_percent=80 \
  checkpointing.load=checkpoints/model.ckpt

The base anysplat model is downloaded automatically from Hugging Face (lhjiang/anysplat). Our fine-tuned checkpoints is provided here:

If you find this work useful, please cite our paper (and AnySplat):

@misc{singhal2026adaptivesplattextureawarecontrollable3d,
      title={AdaptiveSplat:Texture Aware Controllable 3D Gaussian Allocation for Feed-Forward Reconstruction}, 
      author={Badrinath Singhal and Srihari K G and Sreehari Iyer and Ankit Dhiman and Venkatesh Babu Radhakrishnan},
      year={2026},
      eprint={2607.04256},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.04256}, 
}

This project is built on top of AnySplat, and reuses components from DUST3R / CroCo, VGGT, and gsplat. We thank the authors for releasing their code.

Released under the MIT License.