CodeFormer: Blind face restoration with transformer codebook lookup
Transformer-based face restoration using vector-quantized codebook lookup.
Learn more about CodeFormer
from basicsr.archs.codeformer_arch import CodeFormer\nnet = CodeFormer().cuda()\noutput = net(input_image, w=0.5, fidelity_weight=0.5)
Codebook-based restoration
Uses vector quantized codebook lookup to map degraded facial regions to learned high-quality representations, rather than direct pixel-level regression. This approach constrains outputs to realistic face distributions learned during training.
Blind restoration capability
Operates without explicit knowledge of degradation type or severity, handling mixed degradation scenarios including compression artifacts, noise, blur, and missing regions. The model generalizes across different degradation patterns without task-specific fine-tuning.
Multi-task framework
Supports face restoration, inpainting, and colorization through a unified architecture with task-specific checkpoints. Includes video processing support for temporal consistency and integration with multiple face detection backends including dlib and RetinaFace.
from basicsr.archs.codeformer_arch import CodeFormer
import torch
net = CodeFormer().cuda()
net.eval()
with torch.no_grad():
restored_face = net(input_face_tensor, w=0.5, fidelity_weight=0.5)
result = restored_face[0].cpu().clamp(0, 1)Initial release provides pre-trained models; release notes do not specify breaking changes, requirements, or setup steps.
- –Download pre-trained models from this release for CodeFormer inference.
- –Release notes lack migration details, dependency versions, and configuration requirements.
Related Repositories
Discover similar tools and frameworks used by developers
faiss
Efficient approximate nearest neighbor search for billion-scale vectors.
deepface
Python library wrapping multiple face recognition deep learning models.
whisper.cpp
Lightweight, cross-platform speech recognition engine delivering real-time transcription with minimal dependencies and optimized inference performance.
lerobot
PyTorch library for robot imitation learning and sim-to-real transfer.
Mask2Former
Unified transformer architecture for multi-task image segmentation.