GFPGAN
Rebuilds blurry and damaged faces in a single pass.
Uses a pretrained face GAN as a prior to restore facial detail — eyes, skin, hair — from low-quality inputs, without slow per-image optimisation. Our ONNX build takes a 512 × 512 face.
- Output
- 512px
- Parameters
- 85M
- CPU time per 512 px face
- 1.3s
- ONNX file
- 340MB
Example output
Drag the handle to compare. Produced by running this model.
Technical details
- Architecture
- U-Net with a StyleGAN2 face prior
- Checkpoint
- v1.4
- Input
- Aligned face crop
- Output
- Restored face, 512 × 512
- ONNX build
- FP32, opset 11, 512 × 512 input
- Benchmark
- ONNX Runtime CPU, Apple M2 Pro
Good for
- Old family photos
- Blurry portraits
- Low-quality video frames
Files in our release
The same model can ship under more than one name. Check the SHA-256 to confirm a download.
| File | Size | SHA-256 |
|---|---|---|
| GFPGANv1.4.onnx | 340.3 MB | cd7311b8d9e13cdb… |
Related models
All models
Credits and license
- Original model
- GFPGAN
- Created by
- Tencent ARC
- License
- Apache 2.0
- Changes
- Official v1.4 weights converted to an FP32 ONNX file
GFPGAN's license also lists StyleGAN2 code (NVIDIA Source Code License) and DFDNet code (CC BY-NC-SA 4.0), which carry non-commercial terms. Review them before commercial use.
Check the license terms before using this model in a product. More permissive models
Building with AI?
Our models are available as APIs for teams who want to build on them. Tell us what you're working on.