unreflectanything 0.1.6

Deep learning method for removing specular reflections from RGB images.

UnReflectAnything

RGB-Only Highlight Removal by Rendering Synthetic Specular Supervision

UnReflectAnything inputs any RGB image and removes specular highlights, returning a clean diffuse-only outputs. UnReflectAnything works on both natural indoor and surgical/endoscopic domain data.

Installation

pip install unreflectanything

Install UnReflectAnything from PyPI. Python 3.12 is recommended.

For GPU support, make sure PyTorch comes with CUDA version for your system (see PyTorch Get Started).

---

Setting up

Pretrained weights are not included in the package. After installing, download them once:

unreflectanything download-weights

Weights are stored by default in ~/.cache/unreflectanything/weights (or $XDG_CACHE_HOME/unreflectanything/weights if set). Use --output-dir to choose another location. To use a custom weights repository (e.g. on Hugging Face), set the environment variable UNREFLECTANYTHING_WEIGHTS_REPO.

For the download command to work, install the optional dependency: pip install unreflectanything[weights].

---

CLI usage

After installation you get the unreflectanything command (aliases: unreflect, ura; all three do the same thing). For development, use pip install -e . so the same entry points work from the repo.

Subcommand	Description
train	Run training (e.g. unreflectanything train --config config_train.yaml)
test	Run evaluation (unreflectanything test --config config_test.yaml)
inference	Run inference on an image directory (unreflectanything inference --input /images --output /unref_images)
download	Download checkpoint weights
sweep	Launch a Weights & Biases sweep
agent	Run a W&B sweep agent
completion	Print shell completion (bash/zsh): source <(unreflectanything completion bash)

Inference expects a YAML config with weights_path, input_dir, output_dir, and other options (see config_inference.yaml). You can point weights_path to the cache directory after running download-weights.

---

Python API

Use the package programmatically:

import unreflectanything as ura

# Run training or testing
ura.run_pipeline(mode="train")   # or mode="test"

# Run inference from options
from pathlib import Path
from unreflectanything import InferenceOptions, run_inference

options = InferenceOptions(
    weights_path=Path("path/to/full_model_weights.pt"),
    input_dir=Path("input/images"),
    output_dir=Path("output/diffuse"),
)
ura.run_inference(options)

# Utility: compute highlight mask from RGB batch [B,3,H,W]
import torch
mask = ura.compute_highlight_mask(rgb_batch, threshold=0.7)  # [B,1,H,W]

# Default weights cache directory
cache_dir = ura.get_weights_cache_dir()