🧠 Attribute-Leakage-Free Editing (ALE)

Official Implementation of [Addressing Text Embedding Leakage in Diffusion-based Image Editing (ICCV 2025)]

ALE is a training-free image editing framework that prevents attribute leakage in text-guided diffusion-based editing.
It is accompanied by ALE-Bench, a benchmark for evaluating attribute leakage in multi-object text-based image editing.

ALE performs lakage-free edits guided by text prompts, while prior methods show attribute leakages.

---

📁 Project Structure

ale/                     # Core implementation of ALE
│
├── edit.py                  # Main editing pipeline
├── pipeline_ead.py          # Diffusion pipeline for editing
├── ore.py                   # Object-Restricted Embedding (ORE)
├── rgb_cam.py               # Region-Guided Blending for Cross-Attention Masking (RGB-CAM)
├── attention_control.py     # Background Blending (BB) and self-attention injection
├── sam_utils.py             # SAM-based segmentation utilities
├── config.py                # Configuration utilities
│
ale_bench/               # ALE-Bench benchmark dataset
│   ├── images/          # Source images
│   ├── masks/           # Object masks (for evaluation)
│   └── prompts/         # Source–target prompt pairs
│
evaluate/                # Evaluation scripts
│   ├── evaluate_ale.py       # Metric computation & statistics generation
│   └── metrics_calculator.py # Metric definitions (CLIP, PSNR, SSIM, LPIPS, etc.)
│
outputs/                 # SAM outputs
results/                 # Edited outputs generated by ALE
stats/                   # Evaluation results (CSV + XLSX)
│
ale_gradio.py            # Gradio demo interface
run_ale_bench.py         # Run ALE on ALE-Bench
README.md
.gitignore

---

⚙️ Installation

1️⃣ Create the Conda environment

conda env create -f environment.yaml

2️⃣ Install SAM2

Follow the official installation instructions.

3️⃣ Activate the environment

conda activate ale

---

🚀 Usage

1️⃣ Run Gradio Demo

Launch the interactive editing interface:

python ./ale_gradio.py

This opens a Gradio app where you can:

• Load source images and masks
• Enter source and target prompts
• Perform localized, leakage-free text-guided editing

---

2️⃣ Run ALE on ALE-Bench

Automatically perform batch editing on the benchmark dataset:

python ./run_ale_bench.py

This script:

• Loads images, prompts, and masks from ale_bench/
• Runs the ALE pipeline
• Saves generated results under ./results/

---

3️⃣ Evaluate Editing Results

Compute quantitative metrics and aggregate statistics:

python ./evaluate/evaluate_ale.py

This script:

• Loads edited results from ./results/

• Computes metrics (TELS, TILS, Editing Performance, PSNR, SSIM, LPIPS, etc.)

• Saves:

  • evaluation_results.csv
  • Aggregated Excel summaries (ale_mean_*.xlsx)

All saved under the ./stats/ directory.

Example output:

stats/
├── evaluation_results.csv
├── ale_mean_per_target_count.xlsx
├── ale_mean_per_edit_type.xlsx
├── ale_mean_per_target_count_edit_type.xlsx
└── ale_mean_all.xlsx

---

📊 Metrics Overview

Metric	Description	Direction
TELS	Target-External Leakage Score — leakage into background/unrelated regions	Lower is better
TILS	Target-Internal Leakage Score — leakage between different edited objects	Lower is better
Editing Performance	CLIP-based alignment between target prompts and edited regions	Higher is better
PSNR / SSIM / LPIPS / Structure Distance	Standard image similarity metrics	Higher / Lower accordingly

---

🎨 Results

ALE achieves faithful, localized, and leakage-free edits guided by text prompts.

Qualitative comparison with existing editing methods.

🧩 Citation

If you use ALE or ALE-Bench, please cite:

@InProceedings{mun2025ale,
    author    = {Mun, Sunung and Nam, Jinhwan and Cho, Sunghyun and Ok, Jungseul},
    title     = {Addressing Text Embedding Leakage in Diffusion-based Image Editing},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2025},
    url       = {https://arxiv.org/abs/2412.04715},
  }

---

📬 Contact

Sunung Mun POSTECH AI Graduate School 📧 [email protected]

Jinhwan Nam POSTECH AI Graduate School 📧 [email protected]

---

🙇‍♂️ Acknowledgements

This repository builds upon the open-source implementations of InfEdit and Prompt-to-Prompt.

---

💡 ALE: Making diffusion-based image editing faithful, localized, and leakage-free.