IterDiff (ICIP 2025)

IterDiff: Training-Free Iterative Face Editing via Efficient CLIP-guided Memory Bank

Chun-Yao Chiu¹, Feng-Kai Huang², Teng-Fang Hsiao¹, Hong-Han Shuai¹, Wen-Huang Cheng²

¹Institute of Electrical and Computer Engineering, National Yang Ming Chiao Tung University ²Department of Computer Science and Information Engineering, National Taiwan University

---

Project Overview

This repository provides inference and evaluation pipelines for IterDiff and related methods (ip2p, scfg, iterdiff, and emilie). It includes:

• Multi-step face editing inference
• Attention control and memory bank mechanisms
• Quantitative evaluation with CLIP-I, LPIPS, and ImageReward

Environment Setup

Using Conda is recommended (environment.yml is provided, Python 3.12):

conda env create -f environment.yml
conda activate iterdiff

Data Preparation

This project uses the FFHQ download script from StyleGAN2.

cd datasets/ffhq
python download_ffhq.py --images

After downloading, make sure images are located at:

• datasets/ffhq/images1024x1024/

Quick Start

1. Run Inference

   Use script_eval.py to generate multi-step editing results.

   Example (IterDiff):

   python script_eval.py \
    --type iterdiff \
    --mb_size 30 \
    --mb_save_topk 20 \
    --exp_title exp_iterdiff \
    --iter_edit_bench iter_edit_bench_1000.json \
    --results_dir results

2. Compute Metrics

   Use compute_metric.py: