Match-and-Fuse: Consistent Generation from Unstructured Image Sets

Kate Feingold  ·  Omri Kaduri  ·  Tali Dekel

Match-and-Fuse is a zero-shot, training-free method for consistent controlled generation of unstructured image sets — collections that share a common visual element, yet differ in viewpoint, time of capture, and surrounding content.

---

Contents: Method · Installation · Data Setup · Running · Gradio Demo · Evaluation · DINO-MatchSim · Citation

---

Method

Three components plug into the FLUX denoising loop:

Component	What it does
Pairwise Consistency Graph	The image set is modeled as a graph, each edge pairs two images for joint denoising on a shared canvas, activating the grid prior.
Multiview Feature Fusion (MFF)	Each image's features are aggreagred with the features of matched locations in other images, consolidating appearance across the graph.
Feature Guidance	At each denoising step, latents are refined with a gradient-descent step minimizes a feature-level matching objective.

Installation

conda create -n match-and-fuse python=3.10 -y
conda activate match-and-fuse
git clone https://github.com/kate-feingold/match-and-fuse.git
cd match-and-fuse
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124  # adjust to your CUDA version
pip install -r requirements.txt
huggingface-cli login  # for gated FLUX.1 [dev] weights

Download the Depth Anything V2 checkpoint and place it in checkpoints/.

Data Setup

Downloads and resizes images, and computes pairwise matches in one script:

# images + precomputed matches (faster inference)
./scripts/prepare_data.sh --dataset all --matches

--dataset accepts dreambooth, customconcept101, or all.

--matches can be omitted to auto-compute them on the fly during inference.

See data/README.md for more details on data structure.

Running

# General command after `./scripts/prepare_data.sh`
 python main.py --images data/images/<dataset>/<scene>/ --captions_path data/benchmark_captions/<scene>_p<id>.json --save_path results/my_edit

# Read-to-run example
python main.py --images data/images/sample_pet_cat1/ --captions_path data/benchmark_captions/pet_cat1_p1.json --save_path results/my_edit

Or generate captions on the fly with a VLM:

# set a key depending on --caption_model
export OPENAI_API_KEY=sk-...       # gpt-* models
export ANTHROPIC_API_KEY=sk-...    # claude-* models
export GOOGLE_API_KEY=...          # gemini-* models

python main.py --caption_model gpt-4o --prompt_shared "a marble statue" --prompt_theme "autumn" ...

• --images accepts a folder or a list of files (--images img1.png img2.png img3.png)
• --caption_model accepts any OpenAI, Anthropic, or Gemini model. The provided captions can serve as a reference for writing your own, see data/README.md.
• --matches_dir is optional: matches are auto-discovered inside data/, or computed on the fly if not found.
• --flowedit enables inversion-free localized editing (requires source prompts in the captions JSON).

Gradio Demo

A local UI for running the method on data/ scenes or custom images.

python gradio_demo.py
# → http://localhost:5001

Evaluation

python eval.py --dir results/my_edit/

Computes DINO-MatchSim (multi-view consistency) and T2I/LongCLIP-L (alignment with prompts auto-detected from captions.json).

Method	CLIP ↑	DreamSim ↑	DINO-MatchSim ↑
FLUX Kontext	0.65	0.78	0.57
IC-LoRA	0.65	0.71	0.65
FLUX	0.67	0.76	0.66
Edicho	0.65	0.81	0.72
Match-and-Fuse (ours)	0.66	0.85	0.80
w/o Pairwise Graph	0.66	0.82	0.75
w/o MFF	0.66	0.83	0.78
w/o Feat Guidance	0.66	0.82	0.76

DINO-MatchSim

DINO-MatchSim is also available as a standalone pip package — see dino_matchsim/README.md for more details.

pip install dino-matchsim[bg]

Citation

@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}
}

Acknowledgments

We thank the following projects that were used for the codebase:

• x-flux — FLUX and ControlNet
• Depth Anything V2 — depth condition extraction
• RoMa — dense matching implementation