---
datasets:
- depth-anything/DA3-BENCH
base_model:
- zoooooooyuan/VGGT-1B
- depth-anything/DA3-GIANT-1.1
pipeline_tag: image-to-3d
library_name: peft
tags:
- 3D-reconstruction
- test-time-adaptation
- self-supervised-learning
- depth-estimation
- 3d-vision
- peft
---

<div align="center">

<h1 style="text-align:center;">
Free Geometry: Refining 3D Reconstruction from Longer Versions of Itself
</h1>

_Test-time self-evolution for feed-forward 3D reconstruction without 3D ground truth_

<p align="center">
<a href="https://arxiv.org/abs/2604.14048" target="_blank">
<img alt="Paper" src="https://img.shields.io/badge/arXiv-2604.14048-red?logo=arxiv" height="20" />
</a>
<a href="https://github.com/hiteacherIamhumble/Free-Geometry" target="_blank">
<img alt="Code" src="https://img.shields.io/badge/GitHub-Free--Geometry-181717?logo=github" height="20" />
</a>
<a href="https://huggingface.co/PeterDAI/Free-Geometry" target="_blank">
<img alt="Model Weights" src="https://img.shields.io/badge/HuggingFace-Model_Weights-yellow?logo=huggingface" height="20" />
</a>
<a href="./LICENSE" target="_blank">
<img alt="License" src="https://img.shields.io/badge/License-CC_BY_4.0-lightgrey.svg" height="20" />
</a>
</p>

</div>

# Model Description

This repository provides a **LoRA / PEFT adapter** for **Free-Geometry** fine-tuning on top of a pretrained base model (VGGT and DepthAnything3).

This is **not** a full standalone model checkpoint. The repository only contains the adapter weights and configuration needed to apply the fine-tuned update to the original base model.

Use this adapter if you want the Free-Geo fine-tuned behavior while keeping the original base model separate.

# What Is In This Repository?

This repository is expected to contain:

- `adapter_config.json`
- `adapter_model.safetensors`
- `README.md`

These files define the PEFT adapter and tell PEFT how to attach the LoRA weights to the base model.

# Base Model

This adapter was trained on top of:

- [zoooooooyuan/VGGT-1B](https://huggingface.co/zoooooooyuan/VGGT-1B)
- [depth-anything/DA3-GIANT-1.1](https://huggingface.co/depth-anything/DA3-GIANT-1.1)


# Training Setup

This adapter corresponds to:

- Backbone: `DA3` or `VGGT`
- Method: `Free-Geo`
- Dataset: `eth3d`, `7scenes`, `hiroom`, or `scannetpp` (all come from [depth-anything/DA3-BENCH](https://huggingface.co/depth-anything/DA3-BENCH)

## Inference With PEFT

  This repository contains PEFT/LoRA adapter weights. The adapters are not standalone models, so first load the DA3 base model, then attach one adapter from this repo.

  Example using the DA3 + Free-Geo HiRoom adapter:

  ```python
  import torch
  from peft import PeftModel
  from depth_anything_3.api import DepthAnything3

  base_model_id = "depth-anything/DA3-GIANT-1.1"
  adapter_repo_id = "PeterDAI/Free-Geometry"
  adapter_subfolder = "DA3+Free-Geo/hiroom"

  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

  model = DepthAnything3.from_pretrained(base_model_id).to(device)
  model.model.backbone.pretrained = PeftModel.from_pretrained(
      model.model.backbone.pretrained,
      adapter_repo_id,
      subfolder=adapter_subfolder,
      is_trainable=False,
  )
  model.eval()

  prediction = model.inference(
      ["image1.jpg", "image2.jpg"],
      export_dir="output/da3_free_geometry_hiroom",
      export_format="mini_npz",
  )

  print(prediction.depth.shape)
```
  To use another adapter from this repo, keep adapter_repo_id the same and change only adapter_subfolder, for example:

  ```adapter_subfolder = "DA3+Free-Geo/eth3d"```

# Repository Variants

Free Geometry is designed for test-time adaptation, so you can find different model variants for each datasets.

- `DA3-Free-Geo-eth3d`
- `DA3-Free-Geo-7scenes`
- `DA3-Free-Geo-hiroom`
- `DA3-Free-Geo-scannetpp`
- `VGGT-Free-Geo-eth3d`
- `VGGT-Free-Geo-7scenes`
- `VGGT-Free-Geo-hiroom`
- `VGGT-Free-Geo-scannetpp`

# Citation

If you use this model, please cite the corresponding project or paper:

```bibtex
@misc{dai2026freegeometryrefining3d,
title={Free Geometry: Refining 3D Reconstruction from Longer Versions of Itself},
author={Yuhang Dai and Xingyi Yang},
year={2026},
eprint={2604.14048},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.14048},
}
```