| Real-ESRGAN x2plus, optimized Core ML package for WaifuFinder |
| Model weights copyright (c) 2021 Xintao Wang. BSD 3-Clause; see LICENSE. |
|
|
| Original weights: https://github.com/xinntao/Real-ESRGAN (RealESRGAN_x2plus.pth) |
| Reference architecture: https://github.com/XPixelGroup/BasicSR (revision 8d56e3a); |
| the architecture source and its Apache-2.0 licence are included as |
| BasicSR-LICENSE. |
|
|
| WaifuFinder converts these weights to Core ML offline. The conversion fuses the |
| network's first convolution into the following upsampling stage, expresses the |
| network's input and output shapes as enumerated Core ML shapes, and records the |
| native validation result in the model's metadata. The weights themselves are |
| unchanged: manifest.json pins the SHA-256 of the original checkpoint and of the |
| architecture definition. |
|
|
| This package replaces the pre-converted RealESRGAN_x2plus_522_fp16 model from |
| hanxiao/real-esrgan-coreml that earlier builds downloaded. WaifuFinder's version |
| adds validated 256 and 384 pixel inputs, so small tiles spend far less work than |
| padding to 522, and it is the only variant with a recorded original-model |
| parity result. See validation.json. |
|
|
| Inference uses only native Swift and Core ML. Conversion used coremltools and |
| PyTorch offline. |
|
|
