metadata
license: openrail
language:
- en
pipeline_tag: image-to-image
datasets:
- mvp18/gscenes_pretrain
tags:
- diffusion
- image-to-image
- image-to-3d
- 3d-reconstruction
- gaussian-splatting
- pose-free
- sparse-view
- rgbd
base_model:
- stabilityai/stable-diffusion-2
Summary
This repository provides checkpoints used in the Gaussian Scenes pipeline for pose-free, sparse-view scene reconstruction. The weights are stored in Diffusers format and organized as two components:
- UNet — denoising backbone (Diffusers UNet) adapted for our pipeline.
- VAE — variational autoencoder used for latent encoding/decoding.
These checkpoints are intended for research use and model reproducibility.
Usage
For a guide on how to use this model, check out the official repository.
Citation
If you use these checkpoints in your work, please cite the associated paper:
@article{
paul2025gaussian,
title={Gaussian Scenes: Pose-Free Sparse-View Scene Reconstruction using Depth-Enhanced Diffusion Priors},
author={Soumava Paul and Prakhar Kaushik and Alan Yuille},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=yp1CYo6R0r},
note={}
}
The HuggingFace paper page can be found here.