BRIDGE: Background Routing and Isolated Discrete Gating for Coarse-Mask Local Editing

BRIDGE addresses coarse-mask local image editing by separating localization support from geometry generation. It uses BridgePath (Main Path + Subject Path) and a learnable Discrete Geometric Gate for token-level positional-embedding routing.

Overview

• BridgePath: Two-path generation where Main Path preserves background context and Subject Path generates editable content from independent noise
• Discrete Geometric Gate: Token-level PE routing that lets subject tokens borrow background-anchored coordinates near fusion regions or keep subject-centric coordinates for geometry freedom
• Lightweight: 13.31M GateBlock parameters (vs ~1.13B for ControlNet-style branches)

Download

1. Base Model (Required)

BRIDGE is built on top of Qwen-Image-Edit-2511. You must download it first:

# Coming soon: auto-download in inference script
# Model: https://huggingface.co/Qwen/Qwen-Image-Edit-2511

2. BRIDGE Model Weights

Pre-trained BRIDGE weights (STE GateBlocks + LoRA) are available on Hugging Face:

https://huggingface.co/PANDATREE/BRIDGE

Download model.safetensors:

from safetensors.torch import load_file
state = load_file("model.safetensors")

# The checkpoint contains:
# - pipe.ste.* → GateBlocks (Discrete Geometric Gate)
# - lora_*     → LoRA adapters (rank 512)

3. Dataset

The BRIDGE training/evaluation dataset (with global_caption/local_caption) is available on Hugging Face:

https://huggingface.co/datasets/PANDATREE/BRIDGE

Contents:

• dataset_qwen_pe_reversed.json — 42,425 training pairs with captions
• dataset_qwen_pe_top1000_captioned.json — 1,000 evaluation pairs with captions
• fixed_images/ — edited results (guided-filter blended)
• ref_gt_fixed/ / ref_gt_fixed_crop/ — reference ground truth
• fixed_masks/ — editing region masks

Note: target_images/ (original edited outputs from Nano-Banana) are not included in this repo. Please obtain them from Apple Pico-Banana-400K (CC BY-NC-ND 4.0).

Requirements

pip install torch torchvision
pip install -r requirements.txt

Quick Start

Gradio Demo

# Place model.safetensors at:
# train/Qwen-Image-Edit-2511_lora-rank512-cfg/step-28000.safetensors

python apps_demo/app_gradio_multi.py

Inference

See DiffSynth-Studio/examples/qwen_image/model_training/train.py for training, and the evaluation/ scripts for metrics computation.

Training

bash training/Qwen-Image-Edit-2511.sh

Citation

@article{xiong2025bridge,
  title={BRIDGE: Background Routing and Isolated Discrete Gating for Coarse-Mask Local Editing},
  author={Peilin Xiong, Honghui Yuan, Junwen Chen, Keiji Yanai},
  journal={arXiv preprint arXiv:2605.07846},
  year={2025}
}

License

Apache 2.0