GeM-NR: Geometry-Aware Multi-View Editing for Nonrigid Scene Changes

project page · paper


GeM-NR is a fast and flexible training-free pipeline for general multi-view consistent image editing, including edits that drastically change the geometry and appearance of the scene.

The code development was done with python 3.13. Install the following packages to re-create our environment:

pip install -r requirements.txt

# Torch
pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu128

# Depth Anything 3
pip install --no-deps -e third_parties/Depth-Anything-3-dev

# RoMa
pip install --no-deps -e third_parties/RoMa-0.0.1

# GeM-NR
pip install -e .

from gemnr import GemNR

gem_nr = GemNR(resolution=512)
    
# im_pil_list = ...
# edit_text_prompt = ...
# anchor_idx = ...

im_pil_list = [gem_nr.crop_resize(im_pil) for im_pil in im_pil_list]
out_im_pil_list = gem_nr.edit(im_pil_list, edit_text_prompt, anchor_idx)

NOTE:

• you need to provide your Hugging Face token as either environment variable export HF_TOKEN=... or extra argument to the class GemNR(resolution=512, token=...).
• if needed, change the caching paths beforehand: export TORCH_HOME=... and export HF_HOME=....

python run.py

You can also use the same script to run the method on your inputs:

python run.py -i "path_to_img1 ... path_to_imgN" -e "Your edit" -a anchor_image_idx

or

python run.py -i path_to_img_folder -e "Your edit" -a anchor_image_idx

See details in run.py.

You can run multiple edits on different sequences in interactive mode:

python run.py --interactive

Install and setup additional dependencies:

pip install -r requirements_anysplat.txt
cd ./third_parties/AnySplat/src/anysplat/model/encoder/backbone/croco/curope
python setup.py build_ext --inplace

Run e.g.:

python third_parties/AnySplat/run_anysplat.py --image_folder ./results/change_the_horse_statue_to_a_lion_statue_without_a_saddle/edited_sequence 

or check Run AnySplat checkbox in the interactive mode:

If you use this work or find it helpful, please consider citing:

@misc{bengtson2026gemnrgeometryawaremultiviewediting,
      title={GeM-NR: Geometry-Aware Multi-View Editing for Nonrigid Scene Changes}, 
      author={Josef Bengtson and Yaroslava Lochman and Fredrik Kahl},
      year={2026},
      eprint={2606.05142},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.05142}, 
}