Blind image super-resolution using practical degradation model. Includes both 2x and 4x upscaling variants.
- Repository: https://github.com/cszn/BSRGAN
- Paper: Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (ICCV 2021)
- License: Apache-2.0
RRDBNet (Residual-in-Residual Dense Block Network) — 23 RRDB blocks, 64 base features, 32 growth channels. Each RRDB block contains 3 dense blocks with 5 convolutional layers each. Upsampling via nearest-neighbor interpolation followed by convolution (1 step for 2x, 2 steps for 4x).
| Property | model_x2.onnx | model_x4.onnx |
|---|---|---|
| Input | input — float32 [1, 3, H, W] |
input — float32 [1, 3, H, W] |
| Output | output — float32 [1, 3, 2H, 2W] |
output — float32 [1, 3, 4H, 4W] |
| Resolution | Dynamic (any H, W) | Dynamic (any H, W) |
| Normalize | [0, 1] range (divide by 255) | [0, 1] range (divide by 255) |
| Tiling | Yes | Yes |
- Input and output are both RGB images in [0, 1] range.
- Output should be clipped to [0, 1] before converting back to uint8.
- Exported with FP32 precision.
- Architecture is inlined in convert.py (no repo clone needed).
| Property | Value |
|---|---|
| Model license | Apache-2.0 |
| OSAID v1.0 | Open Source AI |
| MOF | Class II (Open Tooling) |
| Training data license | DIV2K (CC0), Flickr2K, WED, OST — standard SR research datasets |
| Training data provenance | Public image restoration benchmarks with synthetic practical degradation (blur, noise, JPEG, resize) |
| Training code | Apache-2.0 |
| Known limitations | Training datasets Flickr2K/WED/OST do not have explicit open-source licenses |
| Published research | Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (ICCV 2021) |
| Inference | Local only, no cloud dependencies |
| Scope | Image upscaling (2x and 4x blind super-resolution) |
| Reproducibility | Full pipeline (setup, convert, clean, demo) |