ddutchie/archwiz-upscalers

ArchWiz Upscalers

Core ML conversions of open image super-resolution and restoration models, used for on-device upscaling in the ArchWiz iOS app. Every file is converted from the original authors' published weights; nothing is retrained or fine-tuned. All credit for the models goes to their authors, listed below with their licences.

Files

File	Model	Scale	Tile	Size	Licence	Original
RealESRGAN_x4plus.mlpackage.zip	Real-ESRGAN x4plus	4x	512 px	31 MB	BSD-3-Clause	Xintao Wang et al.
RealESRGAN_x4plus_anime_6B.mlpackage.zip	Real-ESRGAN x4plus anime 6B	4x	512 px	8 MB	BSD-3-Clause	Xintao Wang et al.
A-ESRGAN_Single.mlpackage.zip	A-ESRGAN (single-scale discriminator, EMA weights)	4x	512 px	31 MB	BSD-3-Clause	Zihao Wei et al.
BSRGAN.mlpackage.zip	BSRGAN	4x	512 px	31 MB	Apache-2.0	Kai Zhang et al.
MMRealSRGAN.mlpackage.zip	MM-RealSR GAN	4x	256 px	49 MB	BSD-3-Clause	Chong Mou et al.
MMRealSRNet.mlpackage.zip	MM-RealSR Net (PSNR-oriented)	4x	256 px	49 MB	BSD-3-Clause	Chong Mou et al.
4xNomos8kSC.mlpackage.zip	4xNomos8kSC	4x	512 px	31 MB	CC-BY-4.0	Philip Hofmann
1xDeNoise_realplksr_otf.mlpackage.zip	1xDeNoise_realplksr_otf	1x	512 px	14 MB	CC-BY-4.0	Philip Hofmann
1xDeJPG_realplksr_otf.mlpackage.zip	1xDeJPG_realplksr_otf	1x	512 px	14 MB	CC-BY-4.0	Philip Hofmann

Each zip contains the .mlpackage, the model's LICENSE.txt and an ATTRIBUTION.txt.

Format

• Core ML ML Program, fp16 weights and compute, iOS 16 / macOS 13 or later.
• Input input: RGB image, fixed square tile (see table). Output output: RGB image, tile × scale.
• Larger images are processed in overlapping tiles and stitched by the caller.
• Converted from PyTorch with spandrel and coremltools. Each conversion is checked against the PyTorch original on a random tile before publishing.

Credits and licences

These models are redistributed under their original licences. Keep the notices in LICENSE.txt / ATTRIBUTION.txt with any copy, and credit the authors if you use them.

Real-ESRGAN x4plus

• File: RealESRGAN_x4plus.mlpackage.zip
• Authors: Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan (Tencent ARC Lab)
• Project: https://github.com/xinntao/Real-ESRGAN
• Paper: Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data
• Licence: BSD-3-Clause, Copyright (c) 2021, Xintao Wang (text)
• Original weights: https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth
• Changes: Converted to Core ML (fp16); weights unchanged.

Real-ESRGAN x4plus anime 6B

• File: RealESRGAN_x4plus_anime_6B.mlpackage.zip
• Authors: Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan (Tencent ARC Lab)
• Project: https://github.com/xinntao/Real-ESRGAN
• Paper: Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data
• Licence: BSD-3-Clause, Copyright (c) 2021, Xintao Wang (text)
• Original weights: https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth
• Changes: Converted to Core ML (fp16); weights unchanged.

A-ESRGAN (single-scale discriminator, EMA weights)

• File: A-ESRGAN_Single.mlpackage.zip
• Authors: Zihao Wei, Yidong Huang, Yuang Chen, Chenhao Zheng, Jingnan Gao
• Project: https://github.com/stroking-fishes-ml-corp/A-ESRGAN
• Paper: A-ESRGAN: Training Real-World Blind Super-Resolution with Attention U-Net Discriminators
• Licence: BSD-3-Clause, Copyright (c) 2021, A-ESRGAN-Team (text)
• Original weights: https://github.com/stroking-fishes-ml-corp/A-ESRGAN/releases/download/v1.0.0/A_ESRGAN_Single.pth
• Changes: Converted to Core ML (fp16); weights unchanged.

BSRGAN

• File: BSRGAN.mlpackage.zip
• Authors: Kai Zhang, Jingyun Liang, Luc Van Gool, Radu Timofte
• Project: https://github.com/cszn/BSRGAN
• Paper: Designing a Practical Degradation Model for Deep Blind Image Super-Resolution
• Licence: Apache-2.0, Kai Zhang, Jingyun Liang, Luc Van Gool, Radu Timofte (text)
• Original weights: https://github.com/cszn/KAIR/releases/download/v1.0/BSRGAN.pth
• Note: Weights from the KAIR v1.0 release (MIT, Copyright (c) 2019 Kai Zhang); BSRGAN itself is Apache-2.0. (text)
• Changes: Converted to Core ML (fp16); weights unchanged.

MM-RealSR GAN

• File: MMRealSRGAN.mlpackage.zip
• Authors: Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, Ying Shan (Tencent ARC Lab)
• Project: https://github.com/TencentARC/MM-RealSR
• Paper: Metric Learning based Interactive Modulation for Real-World Super-Resolution
• Licence: BSD-3-Clause, Copyright (c) 2021, Xintao Wang (text)
• Original weights: https://github.com/TencentARC/MM-RealSR/releases/download/v1.0.0/MMRealSRGAN.pth
• Changes: Converted to Core ML (fp16); weights unchanged.

MM-RealSR Net (PSNR-oriented)

• File: MMRealSRNet.mlpackage.zip
• Authors: Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, Ying Shan (Tencent ARC Lab)
• Project: https://github.com/TencentARC/MM-RealSR
• Paper: Metric Learning based Interactive Modulation for Real-World Super-Resolution
• Licence: BSD-3-Clause, Copyright (c) 2021, Xintao Wang (text)
• Original weights: https://github.com/TencentARC/MM-RealSR/releases/download/v1.0.0/MMRealSRNet.pth
• Changes: Converted to Core ML (fp16); weights unchanged.

4xNomos8kSC

• File: 4xNomos8kSC.mlpackage.zip
• Authors: Philip Hofmann (Phhofm)
• Project: https://huggingface.co/Phips/4xNomos8kSC
• Licence: CC-BY-4.0, Philip Hofmann (text)
• Original weights: https://huggingface.co/Phips/4xNomos8kSC/resolve/main/4xNomos8kSC.safetensors
• Changes: Converted to Core ML (fp16) with a 512px tile input; weights otherwise unchanged.

1xDeNoise_realplksr_otf

• File: 1xDeNoise_realplksr_otf.mlpackage.zip
• Authors: Philip Hofmann (Phhofm)
• Project: https://huggingface.co/Phips/1xDeNoise_realplksr_otf
• Licence: CC-BY-4.0, Philip Hofmann (text)
• Original weights: https://huggingface.co/Phips/1xDeNoise_realplksr_otf/resolve/main/1xDeNoise_realplksr_otf.safetensors
• Changes: Converted to Core ML (fp16) with a 512px tile input; weights otherwise unchanged.

1xDeJPG_realplksr_otf

• File: 1xDeJPG_realplksr_otf.mlpackage.zip
• Authors: Philip Hofmann (Phhofm)
• Project: https://huggingface.co/Phips/1xDeJPG_realplksr_otf
• Licence: CC-BY-4.0, Philip Hofmann (text)
• Original weights: https://huggingface.co/Phips/1xDeJPG_realplksr_otf/resolve/main/1xDeJPG_realplksr_otf.safetensors
• Changes: Converted to Core ML (fp16) with a 512px tile input; weights otherwise unchanged.

Checksums (SHA-256)

334365e076db20f6df0ad8c80ca854443fb61e3c382cc7fb19779d24052211d9  RealESRGAN_x4plus.mlpackage.zip
a1ae41e2217c5a7db5605266b399bb1e0c36b21d4166793cf25968f48e113674  RealESRGAN_x4plus_anime_6B.mlpackage.zip
42c5d631e0510ec6aa035ae198be2f64ffd4027ec92e68ce34319ac37acc3325  A-ESRGAN_Single.mlpackage.zip
5f0f0fcc22324bcc89e5b3e3cad3977eef23af40b49376def53046990230b3f7  BSRGAN.mlpackage.zip
f1b37d8a6843b68cc483940c4ce4131ce6bc766535a77e0a8cec96fd1d59c9ed  MMRealSRGAN.mlpackage.zip
e427f343097a707a3f20fd86c208c58aa417e343c46796b8b812117ab9d6b2de  MMRealSRNet.mlpackage.zip
913ae058685b1d0e706ee5e78eb7d18d1194e0625e45edd0066f7af190883f89  4xNomos8kSC.mlpackage.zip
7ea248d177be89b34e2b8c58169c26cb4c901020b55485104ab7929c6d67f1d4  1xDeNoise_realplksr_otf.mlpackage.zip
c7f94e19f45f4e418d07914b8c77849a77e38bf4aff7523b27c0984bba6ea988  1xDeJPG_realplksr_otf.mlpackage.zip

Disclaimer

Provided as is, without warranty of any kind, as set out in each licence. Not affiliated with or endorsed by the original authors.