EraserDiT: Fast Video Inpainting with Diffusion Transformer Model

---

🗺️ Open-Source Roadmap

🛠️ In Progress

• Gradio demo
• Multi-GPU inference support

✅ Completed

• Single-GPU inference
• Model weights release
• Paper publication

---

🚀 Overview

EraserDiT: Interactively removes specified objects and automatically generates the corresponding prompts. It processes a 2K‑resolution video (2160×2100, 97 frames) in only 65 seconds on a single NVIDIA H800 GPU without any acceleration. Experiments show strong performance in content fidelity, texture restoration, and temporal consistency.

---

🎯 Install dependencies

pip install -r requirements.txt

---

🧸 Inference

EraserDiT requires >60GB GPU memory for a 2K‑resolution video. Multi‑GPU support is in progress and will be open‑sourced later.

python inference.py --vid_path data/10268234.mp4 --mask_path data/10268234_mask.mp4 --prompt "There is a bridge over the lake." 

---

📜 Citation

If you find our work helpful, please consider giving a star 🌟 and citation 📝

@article{liu2025eraserdit,
  title={EraserDiT: Fast Video Inpainting with Diffusion Transformer Model},
  author={Liu, Jie and Hui, Zheng},
  journal={arXiv preprint arXiv:2506.12853},
  year={2025}
}