R-SIE

This repo provides the official code of TACL paper: Overcoming Source Object Grounding for Semantic Image Editing

Official implementation (PyTorch + diffusers) of
“Overcoming Source-Object Grounding for Semantic Image Editing”
accepted to TACL (2025).
We introduce a Region-wise Diffusion Process (RwDP) that cleanly separates background reconstruction from object manipulation, eliminating the grounding errors that plague prior SIE pipelines.

📦 Repository layout

r-sie/
├─ src/                # source codes
│   └─ RwDP_pipeline.py     # huggingface diffusers pipeline for region-wise diffusion process
└─ README.md

Planned additions

• code for data generation process – automatic triplet generation (code & instructions) 🚧
• test_sets/ – links to the cleaned evaluation splits 🚧
• requirements – required packages to run codes 🚧

🔧 Prerequisites

package	tested version
Python	3.9
PyTorch	1.12
diffusers	0.22.0
transformers	4.36.2
accelerate	0.27.2

---

📄 License

Code is released under the MIT License; see LICENSE. Model checkpoints follow the original Stable Diffusion v1.5 CreativeML license.

---

🙋 Questions & Contact

Open an issue or e-mail y970120@snu.ac.kr. We welcome pull requests for bug-fixes or documentation!