Any Resolution Any Geometry: From Multi-View To Multi-Patch

Wenqing Cui¹, Zhenyu Li^1,†, Mykola Lavreniuk^2,†, Jian Shi¹, Ramzi Idoughi¹, Xiangjun Tang¹, Peter Wonka¹.
¹KAUST, ²Space Research Institute NASU-SSAU
^†Equal contribution

✨ NEWS

• 2026-02-27: Initially release
• 2026-02-20: Accepted to CVPR 2026.

Environment Setup

Requirements: Python ≥ 3.10, CUDA 12.4

conda create -n urgt python=3.10 -y
conda activate urgt
pip install -r requirements.txt

Pre-Train Model

Download checkpoints from HuggingFace:

https://huggingface.co/Kingslanding/Any-Resolution-Any-Geometry/tree/main

Checkpoint	Training data	Recommended use
ckpt_best.pth	U4K dataset	U4K benchmark evaluation
ckpt_promask_best.pth	U4K dataset with PRO model masks	Zero-shot evaluation

Place the downloaded checkpoints under work_dir/ckpts/:

work_dir/
└── ckpts/
    ├── ckpt_best.pth
    └── ckpt_promask_best.pth

You can also download directly from the command line:

mkdir -p work_dir/ckpts
huggingface-cli download Kingslanding/Any-Resolution-Any-Geometry \
    ckpt_best.pth ckpt_promask_best.pth \
    --local-dir work_dir/ckpts

User Inference

The inference script runs the full URGT pipeline on a single image:

1. Depth Anything v2 → coarse relative depth
2. Metric3D v2 → coarse surface normals
3. URGT refiner → high-resolution refined depth + normals

Basic usage

python tools/infer.py \
    --image    path/to/image.jpg \
    --checkpoint work_dir/ckpts/ckpt_best.pth \
    --output-dir ./output

With pre-computed coarse predictions

If you already have coarse depth/normal maps (.npy), pass them directly to skip Steps 1 and 2:

python tools/infer.py \
    --image         path/to/image.jpg \
    --checkpoint    work_dir/ckpts/ckpt_best.pth \
    --coarse-depth  path/to/coarse_depth.npy \
    --coarse-normal path/to/coarse_normal.npy \
    --output-dir    ./output

Example (provided test image)

python tools/infer.py \
    --image work_dir/examples/lab_8k.jpg \
    --checkpoint work_dir/ckpts/ckpt_best.pth \
    --output-dir work_dir/examples/output \
    --save-intermediates

All arguments

Argument	Default	Description
--image	(required)	Path to input RGB image (JPG/PNG)
--checkpoint	(required)	Path to URGT checkpoint (.pth)
--output-dir	same dir as image	Directory to save results
--save-intermediates	off	Also save coarse depth/normal visualisations
--dav2-encoder	vitl	Depth Anything v2 encoder: vits / vitb / vitl / vitg
--coarse-depth	None	Pre-computed coarse depth (.npy, shape [H, W]); skips DAv2
--metric3d-model	ViT-Small	Metric3D v2 variant: ViT-Small / ViT-Large / ViT-giant2
--coarse-normal	None	Pre-computed coarse normal (.npy, shape [H, W, 3]); skips Metric3D
--patch-split	8 8	Patch grid N_H N_W; image is resized to be divisible by these values
--min-depth	0.001	Minimum depth value in metres
--max-depth	80.0	Maximum depth value in metres
--device	auto	cuda or cpu (defaults to CUDA when available)

Outputs

For an input named image.jpg, the following files are written to --output-dir:

File	Description
image_depth_pred.png	Colour-mapped refined depth
image_depth_pred.npy	Raw refined depth array, shape [H, W]
image_normal_pred.png	Colour-mapped refined surface normals
image_normal_pred.npy	Raw refined normal array, shape [H, W, 3], range [-1, 1]
image_coarse_depth.png	(with --save-intermediates) Colour-mapped coarse depth
image_coarse_depth.npy	(with --save-intermediates) Raw coarse depth array
image_coarse_normal.png	(with --save-intermediates) Colour-mapped coarse normals
image_coarse_normal.npy	(with --save-intermediates) Raw coarse normal array

Citation

If you find our work useful for your research, please consider citing the paper

@inproceedings{cui2026resolutiongeometrymultiviewmultipatch,
  title={Any Resolution Any Geometry: From Multi-View To Multi-Patch},
  author={Cui, Wenqing and Li, Zhenyu and Lavreniuk, Mykola and Shi, Jian and Idoughi, Ramzi and Tang, Xiangjun and Wonka, Peter},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}