RRLS

KO | EN

gitlite — search

#python #javascript #hacktoberfest #html #rust #android #css #react #vue #machine-learning #json #python3

★ 30 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1>🌉 Bridging the Manifold Gap: Riemannian Residual Line Search for One-Step Image Editing</h1>

<p>
    <a href="#-overview">Overview</a> •
    <a href="#-quick-start">Quick Start</a> •
    <a href="#-demo">Demo</a> •
    <a href="#-pipeline">Pipeline</a> •
    <a href="#-evaluation">Evaluation</a> •
    <a href="#-citation">Citation</a>
  </p>
</div>

## ✨ Overview

RRLS is a source-preserving image editing method. It first produces a baseline edit, then generates a stronger candidate, and finally selects a residual candidate that better balances prompt alignment with source preservation.

## 🚀 Quick Start

### 📦 Dependency

RRLS needs:

* Python 3.12
* PyTorch 2.5.0
* PIE-Bench data, please refer to [PnPInversion](https://github.com/cure-lab/PnPInversion)
* Local Stable Diffusion Turbo [checkpoint](https://huggingface.co/stabilityai/sd-turbo)
* Local CLIP ViT-L/14 [checkpoint](https://huggingface.co/sentence-transformers/clip-ViT-L-14) for evaluation

`DINO ViT-B/8` is downloaded automatically by `torch.hub` during structure-distance evaluation.

### 🛠️ Installation

```bash
pip install -r requirement.txt
```

Put the model weights at the default path or pass `--model-root`:

```text
./sd-turbo/
  unet/
  scheduler/
  text_encoder/
  tokenizer/
  vae/
```

Put PIE-Bench at the default path or pass `--pie-root`:

```text
./pie_bench/
  annotation_images/
  mapping_file.json
```

## 🖥️ Demo

```bash
python app.py
```

The demo uses the same default model layout and loads local examples from `images/` when available.

## 🔁 Pipeline

Run the full export pipeline with:

```bash
bash run_pipeline.sh
```

By default it writes:

* `pie_bench/output/ChordEdit/annotation_images`
* `pie_bench/output/RRLSStrong/annotation_images`
* `pie_bench/output/RRLS/annotation_images`

## 📊 Evaluation

Evaluation is performed on PIE-Bench. It reads source images from `pie_bench/annotation_images`, annotations from `pie_bench/mapping_file.json`, and method outputs from `pie_bench/output/<method>/annotation_images`.

Run the full evaluation with:

```bash
bash run_eval.sh
```

By default it writes results to `pie_bench/eval/`:

* `per_sample_metrics.csv`
* `structure_metrics.csv`
* `summary_metrics.csv`
* `paired_stats.csv`

## 📚 Citation

If you find RRLS useful for your research, please consider citing our paper:

```bibtex
@article{zhu2026rrls,
  title={Bridging the Manifold Gap: Riemannian Residual Line Search for One-Step Image Editing},
  author={Zhu, Yihong and others},
  journal={arXiv preprint arXiv:2606.24844},
  year={2026}
}
```

## 🙏 Acknowledgements

This project builds upon prior work in one-step and training-free image editing. We thank the authors of ChordEdit, PIE-Bench, PnPInversion, Stable Diffusion Turbo, CLIP, and DINO for their valuable resources.

RRLS

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README