unreflectanything 0.3.3

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. We trained UnReflectAnything by synthetizing specularities and supervising in DINOv3 feature space.

UnReflectAnything works on both natural indoor and surgical/endoscopic domain data.

---

Installation

pip install unreflectanything

Install UnReflectAnything as a Python Package.

The minimum required Python version is 3.11, but development and all experiments have been bases on Python 3.12.

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

Setting up

After pip-installing, you can use the unreflectanything CLI command, which is also aliased to unreflect and ura. The three commands are equivalent.

With the CLI you can already download the model weights with

unreflectanything download --weights

and some sample images with

unreflectanything download --images

Weights are stored by default in ~/.cache/unreflectanything/weights (or $XDG_CACHE_HOME/unreflectanything/weights if set ; %LOCALAPPDATA%\unreflectanything for Windows). Use --output-dir to choose another location.

Both the weights and images are stored on the HuggingFace Model Repo.

Enable shell completion

Shell completion is available for the bash and zsh shells. Run

unreflectanything completion bash

and execute the echo ... command that gets printed.

Command Line Interface

Get an overview of the available CLI endpoints with

unreflectanything --help   # alias 'unreflect --help' alias 'ura --help'

Refer to the Wiki to get detailed documentation about each endpoint. We report a summary of the available subcommands. Remember that ura is aliased to the unreflectanything command

Subcommand	Description	Command
inference	Run inference on an image directory	ura inference --input /path/to/images --output /path/to/unref_images
train	Run training	ura train --config config_train.yaml
test	Run evaluation on a trained model	ura test --config config_test.yaml
download	Download checkpoint weights, sample images, notebooks	ura download --weights
verify	Verify weights installation and compatibility, as well as dataset directory structure	ura verify --dataset /path/to/dataset
evaluate	Compute metrics on output data	ura evaluate --output /path/to/unref_images --gt /path/to/groundtruth_images/
completion	Print shell completion (bash/zsh):	ura completion bash
cite	Print shell completion (bash/zsh)	ura cite --bibtex

Python API

The same endpoints above are exposed as a Python API. Refer to the Wiki to get detailed documentation about each endpoint. A few examples are reported below

import unreflectanything as ura
import torch

# Get the model class (e.g. for custom setup or training)
ModelClass = ura.model()

# Get a pretrained model (torch.nn.Module) and run on batched RGB
uramodel = ura.model(pretrained=True)  # uses cached weights; run 'ura download --weights' first
images = torch.rand(2, 3, 448, 448, device="cuda")  # [B, 3, H, W], values in [0, 1]
model_out = uramodel(images)  # [B, 3, H, W] diffuse tensor

# File-based or tensor-based inference (one-shot, no model handle)
ura.inference("input.png", output="output.png")
result = ura.inference(images)  # tensor input returns tensor

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

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

Citation

If you include UnReflectAnything in your pipline or research work, we encourage you cite our work. Get the citation entry with

unreflectanything cite --bibtex

or copy it directly from below

@misc{rota2025unreflectanythingrgbonlyhighlightremoval,
      title={UnReflectAnything: RGB-Only Highlight Removal by Rendering Synthetic Specular Supervision}, 
      author={Alberto Rota and Mert Kiray and Mert Asim Karaoglu and Patrick Ruhkamp and Elena De Momi and Nassir Navab and Benjamin Busam},
      year={2025},
      eprint={2512.09583},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.09583}, 
}