StableNormal: Reducing Diffusion Variance for Stable and Sharp Normal

Chongjie Ye*, Lingteng Qiu*, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu#, Xiaoguang Han#

* Equal contribution
# Corresponding Author

SIGGRAPH Asia 2024 (Journal Track)

We propose StableNormal, which tailors the diffusion priors for monocular normal estimation. Unlike prior diffusion-based works, we focus on enhancing estimation stability by reducing the inherent stochasticity of diffusion models ( i.e. , Stable Diffusion). This enables “Stable-and-Sharp” normal estimation, which outperforms multiple baselines (try Compare), and improves various real-world applications (try Demo).

News

🎉 New Release: StableDelight 🎉

We're excited to announce the release of StableDelight, our latest open-source project focusing on real-time reflection removal from textured surfaces. Check out the StableDelight for more details!

Installation:

Please run following commands to build package:

git clone https://github.com/Stable-X/StableNormal.git
cd StableNormal
pip install -r requirements.txt

or directly build package:

pip install git+https://github.com/Stable-X/StableNormal.git

Usage

To use the StableNormal pipeline, you can instantiate the model and apply it to an image as follows:

import torch
from PIL import Image

# Load an image
input_image = Image.open("path/to/your/image.jpg")

# Create predictor instance
predictor = torch.hub.load("Stable-X/StableNormal", "StableNormal", trust_repo=True)

# Apply the model to the image
normal_image = predictor(input_image)

# Save or display the result
normal_image.save("output/normal_map.png")

Additional Options:

If you need faster inference(10 times faster), use StableNormal_turbo:

predictor = torch.hub.load("Stable-X/StableNormal", "StableNormal_turbo", trust_repo=True)

If Hugging Face is not available from terminal, you could download the pretrained weights to weights dir:

predictor = torch.hub.load("Stable-X/StableNormal", "StableNormal", trust_repo=True, local_cache_dir='./weights')

Compute Metrics:

This section provides guidance on evaluating your normal predictor using the DIODE dataset.

Step 1: Prepare Your Results Folder

First, make sure you have generated a normal map and structured your results folder as shown below:

├── YOUR-FOLDER-NAME 
│   ├── scan_00183_00019_00183_indoors_000_010_gt.png
│   ├── scan_00183_00019_00183_indoors_000_010_init.png
│   ├── scan_00183_00019_00183_indoors_000_010_ref.png
│   ├── scan_00183_00019_00183_indoors_000_010_step0.png
│   ├── scan_00183_00019_00183_indoors_000_010_step1.png
│   ├── scan_00183_00019_00183_indoors_000_010_step2.png
│   ├── scan_00183_00019_00183_indoors_000_010_step3.png

Step 2: Compute Metric Values

Once your results folder is set up, you can compute the metrics for your normal predictions by running the following scripts:

# compute metrics
python ./stablenormal/metrics/compute_metric.py -i ${YOUR-FOLDER-NAME}

# compute variance
python ./stablenormal/metrics/compute_variance.py -i ${YOUR-FOLDER-NAME}

Replace ${YOUR-FOLDER-NAME}; with the actual name of your results folder. Following these steps will allow you to effectively evaluate your normal predictor's performance on the DIODE dataset.

Citation

@article{ye2024stablenormal,
  title={StableNormal: Reducing Diffusion Variance for Stable and Sharp Normal},
  author={Ye, Chongjie and Qiu, Lingteng and Gu, Xiaodong and Zuo, Qi and Wu, Yushuang and Dong, Zilong and Bo, Liefeng and Xiu, Yuliang and Han, Xiaoguang},
  journal={ACM Transactions on Graphics (TOG)},
  year={2024},
  publisher={ACM New York, NY, USA}
}