StableDelight: Revealing Hidden Textures by Removing Specular Reflections

StableDelight is a cutting-edge solution for specular reflection removal from textured surfaces. Building upon the success of StableNormal, which focused on enhancing stability in monocular normal estimation, StableDelight takes this concept further by applying it to the challenging task of reflection removal. The training data include Hypersim, Lumos, and various Specular Highlight Removal datasets from TSHRNet. In addition, we've integrated a multi-scale SSIM loss and random conditional scales technique into our diffusion training process to improve sharpness in one-step diffusion prediction.

Background

StableDelight is inspired by our previous work, StableNormal, which introduced a novel approach to tailoring diffusion priors for monocular normal estimation. The key innovation of StableNormal was its focus on enhancing estimation stability by reducing the inherent stochasticity of diffusion models (such as Stable Diffusion). This resulted in "Stable-and-Sharp" normal estimation that outperformed multiple baselines.

Installation:

Please run following commands to build package:

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

or directly build package:

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

Torch Hub Loader 🚀

To use the StableDelight 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/StableDelight", "StableDelight_turbo", trust_repo=True)

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

# Save or display the result
delight_image.save("output/delight.png")

Gradio interface 🤗

We also provide a Gradio interface for a better experience, just run by:

# For Linux and Windows users (and macOS with Intel??)
python app.py

You can specify the --server_port, --share, --server_name arguments to satisfy your needs!