JinJiang/FADRA-VFR

FADRA: Frequency-Aware Diffusion with Residual Adaptation for Video Face Restoration

Official weights of FADRA. FADRA adapts the frozen Wan2.1-T2V-1.3B video diffusion transformer to blind video face restoration with an LQ pixel-aligned feature fusion layer, LoRA, a Repeated Residual Adaptation Head (RRAH), and a Frequency-Aware Loss (FAL).

Files

File	Size	Description
wan-lora.safetensors	378 MB	LoRA (rank 128) on the DiT attention layers and the expanded 3D patch embedding
patch_embedding.pt	0.4 MB	Expanded input patch embedding for LQ-latent feature fusion
residual_head.pt	145 MB	Repeated Residual Adaptation Head (one DiT block)

These three files are used together with the unmodified official Wan2.1-T2V-1.3B DiT, VAE and UMT5 text encoder, which must be downloaded separately from Wan-AI/Wan2.1-T2V-1.3B.

Training: 15,127 face clips from VFHQ + HDTF, 10k iterations, batch size 32, lr 1e-4, 2× A100.

Usage

Download into the code repository:

huggingface-cli download JinJiang/FADRA-VFR --local-dir checkpoints/FADRA

Then run inference on low-quality face videos (512×512):

python inference.py \
    --control_video_dir ./examples/lq \
    --output_dir ./outputs/examples \
    --width 512 --height 512 --steps 30 --cfg 0.0

See the code repository for installation, base-model download and training instructions.

Results

Dataset	PSNR↑	SSIM↑	LPIPS↓	IDD↓	FVD↓
VFHQ-test	29.95	0.8576	0.2378	0.2912	38.97
CelebV-HQ	28.70	0.8326	0.2421	0.3103	48.32

Citation

@article{jiang2026fadra,
  title   = {FADRA: Frequency-Aware Diffusion with Residual Adaptation for Video Face Restoration},
  author  = {Jiang, Jin and Wang, Jia and Hu, Panwen and Zhao, Weiran and Liao, Shengcai},
  journal = {arXiv preprint arXiv:2607.06389},
  year    = {2026}
}