| --- |
| license: mit |
| tags: |
| - esrgan |
| - realesrgan |
| - ncnn |
| - upscaling |
| - super-resolution |
| - image-processing |
| - computer-vision |
| pipeline_tag: image-to-image |
| library_name: ncnn |
| model-index: |
| - name: Real-ESRGAN-x2plus-NCNN |
| results: |
| - task: |
| type: image-to-image |
| dataset: |
| name: General Image Dataset |
| type: custom |
| metrics: |
| - name: PSNR |
| type: psnr |
| value: "benchmark needed" |
| - name: SSIM |
| type: ssim |
| value: "benchmark needed" |
| --- |
| |
| # Real-ESRGAN x2plus NCNN Model |
|
|
| This repository contains the Real-ESRGAN x2plus model converted to NCNN format for efficient inference, particularly suitable for 2x upscaling of non-anime content. |
|
|
| ## Model Details |
|
|
| - **Model Name**: Real-ESRGAN x2plus |
| - **Purpose**: 2x upscaling of images and videos |
| - **Target**: Non-anime/general photo content (as opposed to anime-style content) |
| - **Scale Factor**: 2x (doubles resolution) |
| - **Format**: NCNN (.param and .bin files) |
|
|
| ## File Contents |
|
|
| - `realesrgan_x2plus.param`: NCNN model architecture definition |
| - `realesrgan_x2plus.bin`: NCNN model weights |
| - `README.md`: This documentation file |
|
|
| ## Original Model Information |
|
|
| This model is based on the original Real-ESRGAN x2plus model, which is designed for general image restoration and super-resolution. It performs well on photographs and realistic images rather than anime-style artwork. |
|
|
| ## Use Cases |
|
|
| - Upscaling low-resolution photos |
| - Enhancing video quality (when used with compatible video processing tools) |
| - General image super-resolution |
| - Photo restoration |
|
|
| ## Compatibility |
|
|
| This model is in NCNN format, making it compatible with: |
| - NCNN inference framework |
| - Video2X for video processing |
| - Any application supporting NCNN models |
|
|
| ## Performance Characteristics |
|
|
| - Optimized for 2x upscaling |
| - Balanced quality and performance |
| - Suitable for both CPU and GPU inference (depending on deployment) |
| - Efficient memory usage compared to original PyTorch format |
|
|
| ## License |
|
|
| The original Real-ESRGAN model was released under the MIT License. Please respect the original license terms when using this converted version. |
|
|
| ## Conversion Notes |
|
|
| This model was converted from the original ONNX format to NCNN format to enable efficient inference in NCNN-compatible applications. The conversion was done using the onnx2ncnn tool from the ncnn framework. |
|
|
| ## How to Use |
|
|
| To use this model with NCNN-compatible applications: |
|
|
| ```bash |
| # Example usage with Video2X |
| video2x --ncnn-param realesrgan_x2plus.param --ncnn-bin realesrgan_x2plus.bin --input input.mp4 --output output.mp4 |
| |
| # Or with other NCNN-based tools that support Real-ESRGAN models |
| ``` |
|
|
| ## Training Data |
|
|
| This model was trained on general image datasets to optimize for realistic photo upscaling. The original training data was not included in this repository due to size constraints. |
|
|
| ## Evaluation |
|
|
| The model has been tested qualitatively on various image types and shows good performance on non-anime content. Quantitative benchmarks (PSNR, SSIM) would need to be calculated separately based on your specific use case. |
