---
license: cc-by-nc-4.0
library_name: gaussianformer
tags:
- 3d-gaussian-splatting
- neural-rendering
- gaussianformer
- pytorch_model_hub_mixin
- model_hub_mixin
datasets:
- ShapeSplats/Objaverse_Splats
base_model:
- microsoft/renderformer-v1-base
---

# GaussianFormer, 256 px training checkpoint

The checkpoint after the two 256 px training stages (Gaussian input head, then the whole model, both from
[RenderFormer](https://huggingface.co/microsoft/renderformer-v1-base)'s weights). It is the starting point of the last two
stages that produce [GaussianFormer](https://huggingface.co/shahafvl/gaussianformer), and is published so those stages
can be reproduced exactly. For rendering, use [shahafvl/gaussianformer](https://huggingface.co/shahafvl/gaussianformer).

**Code:** [github.com/SVLwoof/gaussianformer](https://github.com/SVLwoof/gaussianformer)

```bash
BASE=shahafvl/gaussianformer-256px scripts/train.sh   # stages 3 (256 px recovery) and 4 (512 px)
```

Architecture: RenderFormer's layout with 8 px patches and dense cross-attention; stage 3 switches on the projected 2-D
RoPE, the windowed cross-attention and 4 px patches.

## License

Trained on Objaverse_Splats, a subset of [Objaverse](https://objaverse.allenai.org/) whose terms restrict commercial
use; CC-BY-NC-4.0, research and non-commercial use only.