This repository contains a suite of my high-performance AI models for photorealistic upscaling and restoration. Each model, built on PyTorch architectures including SAFMN, ESRGAN, RealPLKSR and others, has been trained on my own dataset: a library of 11,750+ hand-selected and versatile real-world HD images of exceptional quality and information density. The result of this several months-long dataset curation effort is a set of models that excel at rendering authentic textures and objects (skin, fabric, grit, foliage, etc.) and intelligently handling a wide variety of image degradations. Both the dataset and models are under continuous development, with future releases targeting higher fidelity and generalization. All presented PyTorch models can be used in Chainner.
Software Development in Progress
High-End Video Restoration & Upscaling Software in Development. Topaz -- I'm coming for you.
5.06.2026: PureScale 2.0
These are high-quality upscale models based on exceptionally strong but still lightweight architecture - SAFMN
The architecture file was modified by me to prevent the artifacts that are present in the default architecture from the original paper.
To use these models with Chainner you must replace the default SAFMN architecture file with the attached one (safmn.py).
Assuming you're using a portable Chainner installation, the directory of the SAFMN architecture is here: C:\chaiNNer\python\python\Lib\site-packages\spandrel\architectures\SAFMN__arch
2x_SAFMN_PureScale - GAN model for high-fidelity upscaling of real-world images, handles JPEG compression
2x_SAFMN_PureScale_sharper - GAN model for high-quality upscaling of real-world images with a deblurring component, removes JPEG compression too
4x_SAFMN_PureScale - the ultimate GAN model for high-fidelity 4x upscaling of real-world images; handles JPEG compression as well
safmn.py - modified architecture file to use these models with Chainner or ComfyUI
ONNX models are available as well.
27.09.2025: PureScale 1.0
2x_RealPLKSR_PureScale - an upscaling GAN model based on RealPLKSR architecture designed for high-fidelity upscales of your real-world images (or videos). With regard to degradations, it only handles slight JPEG compression. The result is sharp, high-quality outputs that stay true to the original image. FP16 ONNX model is also attached for usage with NVIDIA TensorRT.
22.08.2025: PureVision 1.0
1x_PureVision - this restoration model is based on ESRGAN architecture, trained to remove various types of compression artifacts: JPEG, WEBP, AVC (H.264), HEVC (H.265) and MPEG-2 and restore missing textures. The model also handles moderate noise and grain. The goal of the model is to make degraded images clear, more detailed and artifact-free. It is also suitable for restoring videos as a preparatory step before upscaling.
2x_PureVision - based on ESRGAN architecture, this model is trained similarly to 1x model, but handles heavier noise and grain and, likewise, is intelligent enough not to remove crucial detail at the same time, for example skin texture. The model has the ability to restore missing textures that were erased by heavy compression. It is designed to be a higher quality and more versatile replacement to the well-known official RealESRGAN_x2plus. Additionally, the model is an excellent upscaler for videos, making the result detailed, polished, artifact and noise-free: it is recommended to use the FP16 ONNX variant of the model alongside NVIDIA TensorRT technology to speed up the upscaling process.
