TweezeEdit

[AAAI 2026] Official Implementation

TweezeEdit: Consistent and Efficient Image Editing with Path Regularization

---

🔍 Overview

TweezeEdit is a training-free image editing framework that improves editing consistency by regularizing source–target denoising paths instead of relying on inversion anchors (the inversion noise of source images) or intrusive model modifications.

---

📄 Paper

• ArXiv: https://arxiv.org/abs/2508.10498

---

🛠 Installation

conda env create -f environment.yml
conda activate TweezeEdit

---

🚀 Running the Code

The main entry point is main.py. The framework supports multiple models with shared and model-specific arguments.

Basic Command

python main.py \
  --model lcm \
  --dataset_config ./dataset.json \
  --base_image_path ./ \
  --out_dir ./results

---

⚙️ Supported Models

1️⃣ LCM (Latent Consistency Model) Model

• Default resolution: 512 × 512
• Default inference steps: 15

Key arguments:

• --is_p2p : Enable P2P editing

• --self_replace_steps: P2P editing param

• --cross_replace_steps: P2P editing param

• --reg_mode

  • 1: Approximate but precise strength control (recommended)
  • 2: Original formulation with less precise control

---

2️⃣ FLUX Model

• Default resolution: 1024 × 1024
• Default inference steps: 28
• Requires a HuggingFace access token (--hf_token or .env)

Example:

python main.py \
  --model flux \
  --hf_token YOUR_HF_TOKEN

---

🧩 Gamma Scheduler

The Gamma Scheduler controls the strength of path regularization at each diffusion step. Proper scheduling enables a balanced trade-off between editing and consistency.

⚠️ Editing performance is ultimately bounded by the generation capability of the base model.

---

Default Gamma Scheduling (LCM)

gamma_step = GammaScheduler.generate(
    length=15,
    mode="step",
    value=1.0,
    k=10
)

---

Custom Tail Scheduling (LCM)

For some examples (e.g., examples/0002.jpg), a custom tail schedule yields better consistency–editing trade-offs:

custom_list = [0.0, 0.8, 0.8, 0.8, 0.0, 0.0]

gamma_step = GammaScheduler.generate(
    length=15,
    mode="custom_tail",
    value=1.0,
    k=9,
    custom_list=custom_list
)

---

📌 Notes

• --skip_step: Skips diffusion steps where gamma = 1.0 to reduce computation (performance impact)
• Output images are automatically resized back to original resolution
• For edits that fail or collapse, consider:
  • Adjusting gamma schedule
  • Using a stronger base model

---

📎 Citation

If you find our work helpful, please cite:

@article{mao2025tweezeedit,
  title={Tweezeedit: Consistent and efficient image editing with path regularization},
  author={Mao, Jianda and Wang, Kaibo and Xiang, Yang and Chen, Kani},
  journal={arXiv preprint arXiv:2508.10498},
  year={2025}
}