DINO-MatchSim

Multi-view consistency metric proposed in "Match-and-Fuse: Consistent Generation from Unstructured Image Sets".

Installation

pip install dino-matchsim[bg]

# Without foreground segmentation
pip install dino-matchsim

Usage

from PIL import Image
from dino_matchsim import dino_matchsim_score, DinoMatchSimCfg, BgCfg

input_images  = [Image.open(p) for p in input_paths]   # before edit
output_images = [Image.open(p) for p in output_paths]  # after edit

results = dino_matchsim_score(
    input_images, output_images,
    cfg=DinoMatchSimCfg(tau=0.6, sim_thresh=0.5),  # optional
    bg_cfg=BgCfg(remove_bg=True),                  # optional
    viz_dir="overlays/",                           # optional: save match overlays
)
print(results["dino_matchsim_output"])  # consistency score in (0, 1]
print(results["dino_matchsim_input"])   # baseline (input upper bound)

See DinoMatchSimCfg and BgCfg for all options.

How it works

1. Computes patch-level DINOv3 features from the input (pre-edit) images.
2. Builds foreground-filtered mutual nearest-neighbour correspondences across all image pairs.
3. Measures feature similarity at those fixed correspondence locations in the output (post-edit) images.
4. Returns exp((S̄ − 1) / τ) where S̄ is the mean cosine similarity — higher is more consistent.

Citation

If you use this metric, please cite:

@inproceedings{matchandfuse2026,
  title={Match-and-Fuse: Consistent Generation from Unstructured Image Sets},
  author={Feingold, Kate and Kaduri, Omri and Dekel, Tali},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}