[ICML 2026] Semantic Granularity Navigation in Image Editing

Liangsi Lu¹, Minzhe Guo¹, Xuhang Chen², Yang Shi^1†
1 Guangdong University of Technology, ² Huizhou University
^† Corresponding author

Overview

This repository provides the standalone public release of Navi-FlowAlign, the FlowAlign-based instantiation of NaviEdit for real-image editing.

NaviEdit is a training-free controller that decouples edit progress from model scale traversal. This release keeps only the method-specific code needed for:

• the standalone Navi-FlowAlign editing pipeline
• a lightweight Gradio demo
• PIE-Bench export in the layout expected by the official PnP evaluation code

Resources

• Stable Diffusion 3 Medium (diffusers): stabilityai/stable-diffusion-3-medium-diffusers
• PIE-Bench / PnP evaluation code: cure-lab/PnPInversion

Please make sure you have accepted the Stability AI model license on Hugging Face before downloading the SD3 weights.

1. Environment

• Python 3.12 or newer
• PyTorch with CUDA support
• Stable Diffusion 3 Medium in diffusers format

--model-root can be either:

• a local checkpoint directory such as /path/to/stable-diffusion-3-medium-diffusers
• a Hugging Face model id such as stabilityai/stable-diffusion-3-medium-diffusers

2. Install Dependencies

pip install -r requirement.txt

3. Run the Web Demo

Launch the interactive demo:

python app.py \
  --model-root /path/to/stable-diffusion-3-medium-diffusers \
  --server-port 7860

If you want to reuse defaults from an existing editor config, you can point --config at a YAML file containing method_config::

python app.py \
  --config /path/to/piebench_full.yaml \
  --model-root /path/to/stable-diffusion-3-medium-diffusers \
  --server-port 7860

Running python app.py launches a local Gradio app for Navi-FlowAlign.

• Left panel: upload the source image and provide source / target prompts
• Right panel: view the edited output image
• Bottom section: click built-in examples to auto-fill the inputs

Optional examples can be placed under images/ using the same lightweight demo layout:

images/
  001/
    i.jpg
    meta.jsonl

Each line in meta.jsonl should contain at least:

{"original_prompt": "...", "edited_prompt": "..."}

4. Run PIE-Bench Export

Export predictions in PIE-Bench format with:

python run_pie_bench.py \
  --model-root /path/to/stable-diffusion-3-medium-diffusers \
  --pie-root /path/to/pie_bench

You can also reuse an existing config:

python run_pie_bench.py \
  --config /path/to/piebench_full.yaml \
  --model-root /path/to/stable-diffusion-3-medium-diffusers \
  --pie-root /path/to/pie_bench

--pie-root should point to a PIE-Bench folder containing at least:

1. annotation_images/ — the original PIE-Bench images
2. mapping_file.json — prompt and metadata annotations

Example layout:

pie_bench/
  annotation_images/
  mapping_file.json

The script writes predictions to:

<pie-root>/output/NaviEdit/annotation_images/...

and can optionally synchronize:

<pie-root>/data/mapping_file.json
<pie-root>/data/annotation_images/...

This matches the export pattern expected by the official PIE-Bench / PnP evaluation code.

For PIE-Bench data preparation and evaluation protocol details, please refer to: https://github.com/cure-lab/PnPInversion

5. Repository Structure

naviedit/
  app.py                 # Gradio demo entrypoint
  run_pie_bench.py       # PIE-Bench export entrypoint
  pipeline_factory.py    # Pipeline construction and runtime defaults
  pipeline_naviedit.py   # Standalone Navi-FlowAlign pipeline wrapper
  model.py               # Navi-FlowAlign controller core
  sd3_bundle.py          # SD3 loading, text encoding, VAE adapter
  utils.py               # Config loading and example discovery
  images/                # Optional local demo examples

6. Citation

If you find our work helpful, please consider starring this repository and citing:

@article{lu2026naviedit,
  title={Semantic granularity navigation in image editing},
  author={Lu, Liangsi and Guo, Minzhe and Chen, Xuhang and Shi, Yang},
  journal={arXiv preprint arXiv:2605.21190},
  year={2026}
}