Re-SISR / renarchi

Super-resolution models

Re-SISR

Models

03
01 AdoreLatest release Real-time upscaling for 1080p anime Scale: 2× Format: ONNX

Notes

A successor to Fallin, trained for real-time 1080p anime upscaling. Available in FP16 and FP32, with DirectML and slimmed ONNX variants in the release.

Known limitation: shuffle patterns can appear in dark reds and bright blues.

License
CC BY-NC-SA 4.0

Download weights

FP16 / ONNX2.9 MB — Download Adore FP16 FP32 / ONNX5.7 MB — Download Adore FP32 Release notes & all 8 files
02 FallinLegacy Anime upscaling, in Soft and Strong variants Scale: 2× Format: ONNX

Notes

Two approaches to anime upscaling. Soft focuses on denoising and compression removal. Strong adds more sharpening.

Soft is the recommended starting point. Strong can introduce oversharpened lines and lighten faces.

License
CC BY 4.0
Status
Legacy release · succeeded by Adore

Download weights

Soft / FP162.9 MB — Download Fallin Soft FP16 ONNX Soft / FP325.7 MB — Download Fallin Soft FP32 ONNX Strong / FP162.9 MB — Download Fallin Strong FP16 ONNX Strong / FP325.7 MB — Download Fallin Strong FP32 ONNX Release notes & all files
03 Alchemy General upscaling & compression removal Scale: 4× Format: PTH

Notes

A 4× model for anime, manga covers, novel illustrations, album art and webtoons. Trained on a custom dataset of 3,500 images, mostly hand-picked fanart.

Architecture
RealPLKSR_Dysample
License
CC BY 4.0

Download weights

4x_Alchemy / PTH29.7 MB — Download Alchemy Release notes

Open a model for notes and downloads.

All releases on GitHub

About

I'm renarchi. I train super-resolution models.

Find me on GitHub