HDRFace

KO | EN

gitlite — search

#mysql #svg #terminal #agent #animation #chatbot #rust #tensorflow #tailwindcss #graphql

(★ 9)

HDRFace: Rethinking Face Restoration with High-Dimensional Representation

파일 탐색기

ZIP 다운로드

- 0.yaml
- 01.yaml
- 0123.yaml
- 1.yaml
- 2.yaml
- 23.yaml
- 3.yaml
- 4.yaml
- 45.yaml
- 4567.yaml
- 5.yaml
- 6.yaml
- 67.yaml
- 7.yaml
- - adaface_demo5.gif
  - ms1mv2_train_subset_index.txt
- - augmenter.py
  - five_validation_dataset.py
  - image_folder_dataset.py
  - record_dataset.py
- - align.py
  - mtcnn.py
  - OSDFace
- - run_ir101_ms1mv2.sh
  - run_ir50_ms1mv2.sh
- - eval_5valsets.sh
- - data_utils.py
  - eval_ijbs.sh
  - evaluate_helper.py
  - README.md
  - tinyface_helper.py
  - validate_IJB_S.py
  - validate_tinyface.py
- - eval_ijb.sh
  - README.md
  - validate_IJB_BC.py
- config.py
- convert.py
- data.py
- evaluate_utils.py
- head.py
- inference.py
- LICENSE
- main.py
- net.py
- README.md
- README_TRAIN.md
- requirements.txt
- test.py
- train_val.py
- utils.py
- iresnet.py
- test.py
- - __init__.py
  - arch_util.py
  - basicvsr_arch.py
  - basicvsrpp_arch.py
  - dfdnet_arch.py
  - dfdnet_util.py
  - discriminator_arch.py
  - duf_arch.py
  - ecbsr_arch.py
  - edsr_arch.py
  - edvr_arch.py
  - hifacegan_arch.py
  - hifacegan_util.py
  - inception.py
  - rcan_arch.py
  - ridnet_arch.py
  - rrdbnet_arch.py
  - spynet_arch.py
  - srresnet_arch.py
  - srvgg_arch.py
  - stylegan2_arch.py
  - stylegan2_bilinear_arch.py
  - swinir_arch.py
  - tof_arch.py
  - vgg_arch.py
- - __init__.py
  - data_sampler.py
  - data_util.py
  - degradations.py
  - ffhq_dataset.py
  - paired_image_dataset.py
  - prefetch_dataloader.py
  - realesrgan_dataset.py
  - realesrgan_paired_dataset.py
  - reds_dataset.py
  - single_image_dataset.py
  - transforms.py
  - video_test_dataset.py
  - vimeo90k_dataset.py
- - __init__.py
  - basic_loss.py
  - gan_loss.py
  - loss_util.py
  - losses.py
- - __init__.py
  - fid.py
  - metric_util.py
  - niqe.py
  - niqe_pris_params.npz
  - psnr_ssim.py
  - README_CN.md
- - __init__.py
  - base_model.py
  - edvr_model.py
  - esrgan_model.py
  - hifacegan_model.py
  - lr_scheduler.py
  - realesrgan_model.py
  - realesrnet_model.py
  - sr_model.py
  - srgan_model.py
  - stylegan2_model.py
  - swinir_model.py
  - video_base_model.py
  - video_gan_model.py
  - video_recurrent_gan_model.py
  - video_recurrent_model.py
- - __init__.py
- - __init__.py
  - color_util.py
  - diffjpeg.py
  - dist_util.py
  - download_util.py
  - face_util.py
  - file_client.py
  - flow_util.py
  - img_process_util.py
  - img_util.py
  - lmdb_util.py
  - logger.py
  - matlab_functions.py
  - misc.py
  - options.py
  - plot_util.py
  - registry.py
- __init__.py
- test.py
- train.py
- VERSION
- version.py
- test.py
- vqvae_v2.py
- vqvae_v3.py
- - 00000000.png
  - 00000001.png
  - 00000002.png
  - 00000003.png
  - 00000004.png
  - 00000005.png
  - 00000006.png
  - 00000007.png
  - 00000008.png
  - 00000009.png
- - 00000000.png
  - 00000001.png
  - 00000002.png
  - 00000003.png
  - 00000004.png
  - 00000005.png
  - 00000006.png
  - 00000007.png
  - 00000008.png
  - 00000009.png
- - 00000000.png
  - 00000001.png
  - 00000002.png
  - 00000003.png
  - 00000004.png
  - 00000005.png
  - 00000006.png
  - 00000007.png
  - 00000008.png
  - 00000009.png
- add_usm.py
- codeformer.py
- degradation.py
- face_dataset.py
- ffhq_degradation_dataset.py
- ffhq_degradation_dataset_group.py
- ffhq_degradation_dataset_pair.py
- file_backend.py
- utils.py
- - __init__.py
  - model_config.py
- - __init__.py
  - controlnet_unit.py
  - processors.py
- - __init__.py
  - simple_text_image.py
  - video.py
- - __init__.py
  - xdit_context_parallel.py
- - __init__.py
- - __init__.py
  - flux_lora.py
- - __init__.py
  - attention.py
  - cog_dit.py
  - cog_vae.py
  - downloader.py
  - flux_controlnet.py
  - flux_dit.py
  - flux_infiniteyou.py
  - flux_ipadapter.py
  - flux_lora_encoder.py
  - flux_text_encoder.py
  - flux_vae.py
  - flux_value_control.py
  - hunyuan_dit.py
  - hunyuan_dit_text_encoder.py
  - hunyuan_video_dit.py
  - hunyuan_video_text_encoder.py
  - hunyuan_video_vae_decoder.py
  - hunyuan_video_vae_encoder.py
  - kolors_text_encoder.py
  - lora.py
  - model_manager.py
  - nexus_gen.py
  - nexus_gen_ar_model.py
  - nexus_gen_projector.py
  - omnigen.py
  - qwen_image_dit.py
  - qwen_image_text_encoder.py
  - qwen_image_vae.py
  - qwenvl.py
  - sd3_dit.py
  - sd3_text_encoder.py
  - sd3_vae_decoder.py
  - sd3_vae_encoder.py
  - sd_controlnet.py
  - sd_ipadapter.py
  - sd_motion.py
  - sd_text_encoder.py
  - sd_unet.py
  - sd_vae_decoder.py
  - sd_vae_encoder.py
  - sdxl_controlnet.py
  - sdxl_ipadapter.py
  - sdxl_motion.py
  - sdxl_text_encoder.py
  - sdxl_unet.py
  - sdxl_vae_decoder.py
  - sdxl_vae_encoder.py
  - step1x_connector.py
  - stepvideo_dit.py
  - stepvideo_text_encoder.py
  - stepvideo_vae.py
  - svd_image_encoder.py
  - svd_unet.py
  - svd_vae_decoder.py
  - svd_vae_encoder.py
  - tiler.py
  - utils.py
  - wan_video_camera_controller.py
  - wan_video_dit.py
  - wan_video_image_encoder.py
  - wan_video_motion_controller.py
  - wan_video_text_encoder.py
  - wan_video_vace.py
  - wan_video_vae.py
- - __init__.py
  - base.py
  - cog_video.py
  - dancer.py
  - flux_image.py
  - flux_image_new.py
  - hunyuan_image.py
  - hunyuan_video.py
  - omnigen_image.py
  - pipeline_runner.py
  - qwen_image.py
  - sd3_image.py
  - sd_image.py
  - sd_video.py
  - sdxl_image.py
  - sdxl_video.py
  - step_video.py
  - svd_video.py
  - wan_video.py
  - wan_video_new.py
- - __init__.py
  - base.py
  - FastBlend.py
  - PILEditor.py
  - RIFE.py
  - sequencial_processor.py
- - __init__.py
  - base_prompter.py
  - cog_prompter.py
  - flux_prompter.py
  - hunyuan_dit_prompter.py
  - hunyuan_video_prompter.py
  - kolors_prompter.py
  - omnigen_prompter.py
  - omost.py
  - prompt_refiners.py
  - sd3_prompter.py
  - sd_prompter.py
  - sdxl_prompter.py
  - stepvideo_prompter.py
  - wan_prompter.py
- - __init__.py
  - continuous_ode.py
  - ddim.py
  - flow_match.py
- - __init__.py
- - __init__.py
  - text_to_image.py
  - utils.py
- - __init__.py
- - __init__.py
  - gradient_checkpointing.py
  - layers.py
- __init__.py
- - README.md
  - test.py
  - train.py
- - __init__.py
  - demo_util.py
  - test.py
  - train.py
  - version.py
- eval.py
- eval.sh
- firstimg.png
- pipeline.png
- __init__.py
- inference.py
- vqvae.py
- - accelerate_config_zero2offload.yaml
  - Qwen-Image.sh
- - Qwen-Image.sh
- - Qwen-Image.py
- - Qwen-Image.py
- train.py
- __init__.py
- autoencoder_kl.py
- autoencoder_kl_ose.py
- DCGAN.py
- gate.py
- sd_guidance.py
- sd_unet_forward.py
- sd_unet_forward_official.py
- sd_unifined_model.py
- unet_2d_condition_official.py
- unet_2d_condition_osediff.py
- vae.py
- f3_640.yaml
- restoreformer_usm.yaml
- restoreformer_usm_lq.yaml
- test_qwen.sh
- test_sd.sh
- common.py
- devices.py
- dinov3_fusion_module.py
- discriminator.py
- fusion.py
- ganloss.py
- others.py
- vaehook.py
- vaehook_v2.py
- wavelet_color_fix.py
base.py
generator.py
inference_qwen.py
inference_sd.py
LICENSE
README.md
requirements.txt

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">HDRFace: Rethinking Face Restoration with High-Dimensional Representation</h1>
<div align="center">
<hr>

Zirui Wang<sup>1,2*</sup>&nbsp; Xianhui Lin<sup>2</sup>&nbsp; Yi Dong<sup>2</sup>&nbsp; Bo Wei<sup>2</sup>&nbsp;  Gangjian Zhang<sup>2</sup>&nbsp; Siteng Ma<sup>2</sup>&nbsp; Zebiao Zheng<sup>2</sup>&nbsp; Xing Liu<sup>2</sup>&nbsp; Hong Gu<sup>2</sup>&nbsp; Minjing Dong<sup>1,†</sup>&nbsp;

<sup>1</sup> City University of Hong Kong &nbsp;&nbsp;<sup>2</sup> vivo BlueImage Lab&nbsp;&nbsp;

<sup>*</sup> Work done when interning at vivo. <sup>†</sup> Corresponding authors.

<h4>

<a href="https://arxiv.org/abs/2605.14821">📄 arXiv Paper</a> &nbsp; 
<a href="https://zirui0625.github.io/projects/HDRFace">🌐 Project Page</a> &nbsp; 
<a href="https://huggingface.co/zirui0625/HDRFace">🤗 Hugging Face Models</a>
</h4>

</div>

<div align="center">
<img src="./figures/firstimg.png" width="90%">
</div>
Figure 1. Motivation and improvements of HDRFace. (Left) Existing diffusion-based methods rely solely on low-quality inputs, yielding outputs that lack fine-grained detail and faithful identity preservation. We address this by injecting high-dimensional visual representations into the generative pipeline. (Right) Our method consistently outperforms OSDFace in perceptual quality (LPIPS), identity consistency (ArcFace degree), and image quality (TOPIQ).

## ✨ Highlights

🎯 **High-Dimensional Priors** We propose a high-dimensional representation conditioned face restoration framework that injects DINOv3 semantic features into the conditional branch to provide priors beyond low-quality inputs and ease the ill-posedness caused by missing information.

🚀 **Adaptive Multi-Source Fusion** We design an architecture-independent module SDFM that adaptively fuses low-quality inputs with high-dimensional features, balancing structural consistency and detail fidelity without changing the generative backbone.

⭐️ **Backbone-Agnostic Generalization** Extensive experiments on SD V2.1-base and Qwen-Image demonstrate strong generalization and architecture independence with consistent restoration quality gains.

## 🛠️ 1. Environment Setup

The code is developed using **Python 3.10** and **PyTorch**.

```bash
conda create -n HDRFace python=3.10
conda activate HDRFace

# Install dependencies
pip install -r requirements.txt
```

## 📦 2. Pretrained Weights

Please download the following weights and place them in the `./preset/models` directory.

## 📂 3. Dataset Preparation

### 🦅 Test Datasets
-  We evaluate our method on three datasets: CelebA-Test, LFW-Test, and CelebChild. The download links can be found in  Google Drive ([Link](https://drive.google.com/file/d/1UpjpcQGpp_P4hgOlA4FRLfOBb3V7vV3Y/view?usp=drive_link)) or VQFR ([Link](https://github.com/TencentARC/VQFR)).
## 🚀 4. Usage

### 🎨 Quick Start (Demo)
You can download our restored image for testing here ([Link](https://drive.google.com/file/d/1UpjpcQGpp_P4hgOlA4FRLfOBb3V7vV3Y/view?usp=drive_link)).
Download our pretrained models ([Link](https://huggingface.co/zirui0625/HDRFace)) and place them in `./pretrained` and `./pretrained_qwen`.
```bash
# Test sd
bash ./test_scripts/test_sd.sh
# Test qwen
bash ./test_scripts/test_qwen.sh
```
### 🧪 Evaluation
```bash
bash ./eval/eval.sh
```

## 🙏 Acknowledgements

Our codes are based on OSDFace ([Link](https://github.com/jkwang28/OSDFace)), ODTSR ([Link](https://github.com/RedMediaTech/ODTSR)), thanks for their contribution.

## 📜 Citation

If you find our work or code useful for your research, please cite:

```latex
@misc{wang2026hdrfacerethinkingfacerestoration,
      title={HDRFace: Rethinking Face Restoration with High-Dimensional Representation}, 
      author={Zirui Wang and Xianhui Lin and Yi Dong and Bo Wei and Gangjian Zhang and Siteng Ma and Zebiao Zheng and Xing Liu and Hong Gu and Minjing Dong},
      year={2026},
      eprint={2605.14821},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.14821}, 
}
```

HDRFace

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

HDRFace

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README