BRIDGE Model Weights

BRIDGE uses a main generation path, an independently denoised subject path, and learned discrete positional-encoding routing. This repository now contains both the original Qwen release and two FLUX.2 Klein 9B variants. They are different architectures and their weights are not interchangeable.

Available checkpoints

Release	Path	Required base model	Format
Original Qwen BRIDGE	model.safetensors	Qwen/Qwen-Image-Edit-2511	GateBlocks + rank-512 LoRA
FLUX Sparse / Mask	flux2-klein-9b/sparse-mask/checkpoint-1400	black-forest-labs/FLUX.2-klein-base-9B	Full BF16 transformer + gates, ScheduleFree eval view
FLUX Dense / BBox	flux2-klein-9b/dense-bbox/checkpoint-1400	black-forest-labs/FLUX.2-klein-base-9B	Full BF16 transformer + gates, ScheduleFree eval view

The original Qwen weight file is retained unchanged. The FLUX checkpoints are the two step-1400 eval exports used by the project's comparison Gradio launcher, not the older training-view exports. Each has approximately 18.2 GB of weights. VAE, text encoder, tokenizer and scheduler are obtained separately from the base model; this is not a complete standalone Diffusers pipeline.

FLUX download and usage

hf download PANDATREE/BRIDGE \
  --include 'flux2-klein-9b/*' --local-dir ./bridge-weights

Use the custom BRIDGE transformer and Gradio backend, not the standard base transformer alone. See the FLUX code and instructions and the FLUX release notes. Generation uses exactly 50 inference steps.

Original Qwen usage

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

• pipe.ste.*: GateBlocks (Discrete Geometric Gate).
• lora_*: LoRA adapters (rank 512).

Base model: Qwen/Qwen-Image-Edit-2511. See the BRIDGE repository for the original Qwen inference and training code.

Licensing

This repository contains differently licensed artifacts; the old Apache label must not be applied to the new FLUX weights.

• The original Qwen BRIDGE release retains its Apache-2.0 designation.
• The FLUX weights are derivatives of FLUX.2 Klein base 9B and are distributed under the upstream FLUX Non-Commercial License v2.1. Read the attribution and modification notice. Rights to use the underlying FLUX model are granted directly by Black Forest Labs under that license. These derivatives are not an official or endorsed Black Forest Labs release.

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}
}