BFR_RL — LRPO weights
Model weights for "Beyond the Boundary: RL-Driven Solution Space Exploration for Blind Face Restoration" (LRPO, ECCV 2026).
Code & full instructions: https://github.com/01NeuralNinja/BFR_RL
Files
| file | description |
|---|---|
lrpo_fidelity.pth |
LRPO ControlNet — paper model, fidelity-oriented (ours) |
lrpo_perception.pth |
LRPO ControlNet — perception-oriented, no GT-likelihood regularization (ours) |
v1_face.pth |
DiffBIR-v1 face ControlNet (base) |
face_swinir_v1.ckpt |
DiffBIR SwinIR-face stage-1 (base) |
v2-1_512-ema-pruned.ckpt |
Stable Diffusion 2.1-base |
FaceRewardModel.pth |
face reward model (training only) |
ffhq_captions.tar.gz |
FFHQ per-image captions (training only) |
The two lrpo_*.pth files are our released models; the rest are base/auxiliary weights bundled here for one-stop convenience and remain subject to their original licenses (Stable Diffusion 2.1, DiffBIR).
Download
pip install -U "huggingface_hub[cli]"
hf download wubin1928/BFR_RL --local-dir ./weights
or in Python:
from huggingface_hub import snapshot_download
snapshot_download("wubin1928/BFR_RL", local_dir="./weights")
Usage
See the GitHub repository for inference and training instructions.
License
Code and our released LRPO weights: Apache 2.0. Bundled base weights follow their original licenses.