Real-ESRGAN MLX (Swift)
PyTorch → MLX Swift converted weights for Real-ESRGAN.
Use with mlx-realesrgan-swift.
Files
| File | Variant | Architecture | Size |
|---|---|---|---|
realesr-general-x4v3.safetensors |
SRVGGNetCompact ×4 (general) | 34 conv + 33 PReLU + PixelShuffle | 4.9 MB |
realesr-animevideov3.safetensors |
SRVGGNetCompact ×4 (anime video) | 18 conv + 17 PReLU + PixelShuffle | 2.5 MB |
RealESRGAN_x4plus.safetensors |
RRDBNet ×4 (flagship) | 351 conv across 23 RRDB blocks | 66.9 MB |
RealESRGAN_x4plus_anime_6B.safetensors |
RRDBNet ×4 (anime stills) | 96 conv across 6 RRDB blocks | 17.9 MB |
Conversion details
- Conv2d weights permuted OIHW → OHWI (MLX NHWC convention)
- All tensors float32
License
This conversion (this README, the .safetensors packaging, and any added metadata) is released under the MIT License — see the LICENSE in the companion GitHub repo.
The underlying neural network weights are derived from the official Real-ESRGAN PyTorch releases by Xintao Wang and the BasicSR / XPixelGroup contributors, originally distributed under the BSD 3-Clause License. That license and the original copyright notice continue to apply to the substantive weight content.
Upstream: https://github.com/xinntao/Real-ESRGAN Upstream license: https://github.com/xinntao/Real-ESRGAN/blob/master/LICENSE
Acknowledgements
- xinntao/Real-ESRGAN — original PyTorch implementation and trained weights
- XPixelGroup/BasicSR — architecture definitions
Hardware compatibility
Log In to add your hardware
Quantized