The best place to find AI Upscaling models

A 1x model for Restoration, Denoise . Link to Github Release # 1xDeNoise_realplksr_otf Scale: 1 Architecture: RealPLKSR Architecture Option: realplksr Author: Philip Hofmann License: CC-BY-0.4 Subject: Photography Input Type: Images Release Date: 08.08.2024 Dataset: nomosv2 Dataset Size: 6000 OTF (on the fly augmentations): Yes Pretrained Model: 1xDeJPG_realplksr_otf Iterations: 200'000 Batch Size: 8 Patch Size: 64 Description: A 1x realplksr model to denoise, trained with the realesrgan-otf pipeline, also handles a bit of jpg compression (if stronger jpg compression handling is needed, 1xDeJPG_realplksr_otf can be used).

A 4x model for Restoration . Link to Github Release # 4xArtFaces_realplksr_dysample Scale: 4 Architecture: RealPLKSR with Dysample Architecture Option: realplksr Author: Philip Hofmann License: CC-BY-0.4 Subject: Art Input Type: Images Release Date: 08.08.2024 Dataset: ArtFaces Dataset Size: 5'630 OTF (on the fly augmentations): No Pretrained Model: 4xNomos2_realplksr_dysample Iterations: 139'000 Batch Size: 6 Patch Size: 64 Description: A Dysample RealPLKSR 4x upscaling model for art / painted faces. Based on my ArtFaces Dataset which is a curated version of the metfaces dataset for the purpose of training single image super resolution models.

Restore frames to their original look as on Drawings. Before analog transfer Trained on thousands of 4K HQ Drawing scans, the purpose of this model is restore the frame to its original paper and plastic look, with a particular emphasis on textures and preserving those brushstrokes and natural lines. Not to be used with heavily grainy sources. Note: The random white speckles are an unfortunate side effect of not filtering the dataset.

Upscaling of Numeric Animation from SD to HD with high texture retention. A serious upgrade over 1.0. Much more universality. Can look good on both clean SD and noisy/grainy source frames. It still has a weakness against sources with strong halos and jagged edges and rainbowing. (Recommended that these issues are fixed beforehand with avisynth)

Upscaling of Numeric Animation from SD to HD with high texture retention. Provides the sharpest results yet. About 30% better than the base 2.0 version. But as the title says it's the most aggressive version of NF. Does it's best on animation with little to no textures. The model as well as any version of NF is not suitable at all for overly deteriorated/blurred and low quality TVRips and VHSRips copies of animation pieces.

Upscaling of Numeric Animation from SD to HD with high texture retention. Big leap from 2.0, especially as much as texture retention is concerned. But also deals much better with distant characters/shots.

## AniSD Suite (15 models) **Scale:** 2x (1x for AniSD DB) **Architecture:** SPAN / Compact / SwinIR Small / CRAFT / DAT2 / RPLKSR **Dataset:** Anime frames. Credits to @.kuronoe. and @pwnsweet (EVA dataset) for their contributions to the dataset! **Dataset Size:** ~7,000 - ~13,000 AniSD is a suite of 15 (as of time of writing) specialized SISR models trained to restore and upscale standard definition digital anime from ~2000s and onwards, including both both WEB and DVD releases. Faithfulness to the source and natural-looking output are the guiding principles behind the training of the AniSD models. This means avoiding oversharpened output (which can look especially absurd on standard definition sources), minimizing upscaling artifacts, retaining the natural detail of the source and of course, fixing the standard range of issues found in many DVD/WEB release (chroma issues, compression, haloing/ringing, blur, dotcrawl, banding etc.). Refer to the infographic above for a quick breakdown of the available models, and refer to the Github release for further information.

Link to Github Release Since no official HMA model releases exist yet, I am releasing my hma and hma_medium mssim pretrains. These can be used to speed up and stabilize early training stages when training new hma models. Trained with mssim on nomosv2. ## 4xmssim_hma_pretrain **Scale:** 4 **Architecture:** HMA **Architecture Option:** hma **Author:** Philip Hofmann **License:** CC-BY-0.4 **Purpose:** Pretrained **Subject:** Photography **Input Type:** Images **Release Date:** 19.07.2024 **Dataset:** nomosv2 **Dataset Size:** 6000 **OTF (on the fly augmentations):** No **Pretrained Model:** None (=From Scratch) **Iterations:** 205'000 **Batch Size:** 4 **Patch Size:** 96 **Description:** A pretrain to start hma model training. --- ## 4xmssim_hma_medium_pretrain **Scale:** 4 **Architecture:** HMA **Architecture Option:** hma_medium **Author:** Philip Hofmann **License:** CC-BY-0.4 **Purpose:** Pretrained **Subject:** Photography **Input Type:** Images **Release Date:** 19.07.2024 **Dataset:** nomosv2 **Dataset Size:** 6000 **OTF (on the fly augmentations):** No **Pretrained Model:** None (=From Scratch) **Iterations:** 150'000 **Batch Size:** 4 **Patch Size:** 48 **Description:** A pretrain to start hma_medium model training. --- **Showcase:** slow.pics

Simple pretrain for RealPLKSR Dysample, trained on CC0 content. "Ethical" model

A 1x model for 1x realistic noise degradation model . Github Release Link Name: Ludvae200 License: CC BY 4.0 Author: Philip Hofmann Network: LUD-VAE Scale: 1 Release Date: 25.03.2024 Iterations: 190'000 H_size: 64 n_channels: 3 dataloader_batch_size: 16 H_noise_level: 8 L_noise_level: 3 Dataset: RealLR200 Number of train images: 200 OTF Training: No Pretrained_Model_G: None Description: 1x realistic noise degradation model, trained on the RealLR200 dataset as found released on the SeeSR github repo. Next to the ludvae200.pth model file, I provide a ludvae200.zip file which not only contains the code but also an inference script to run this model on the dataset of your choice. Adapt the ludvae200_inference.py script accordingly by adjusting the file paths at the beginning section, to your input folder, output folder, the folder path holding the ludvae200.pth model, and a folder path where you want the text file to be generated. I made the textfile generation the same way as I did in Kim's Dataset Destroyer, which means you will have each image file logged with each of the values used to degrade that specific image file in the resulting text file, which will append only and never overwrite. You can also adjust the strength settings inside the inference script file to fit to your needs. If you in general want less strong noise for example, you should adjust the temperature upper limit from 0.4 to 0.2 or go even lower. So in line 96 change "temperature_strength = uniform(0.1,0.4)" to "temperature_strength = uniform(0.1,0.2)" just to give an example. These values are defaulted to my needs of my last dataset degradation workflow I used, but feel free to adjust these values. You can also do the same as I did, temporarily using deterministic values with multiple runs to determine the min and max values of noise generation you deem suitable for your dataset needs. See the examples of what this looked like for my last dataset workflow I used my model in.

## Ani4K v2 Ani4K v2, as the successor to the original Ani4K, retains its predecessor's fantastic detail retention, depth of field preservation and faithfulness to the original source. As the name suggests, the model is targeted at modern anime, ranging from high-quality Bluray to crappy WEB releases, for upscaling either to 2K or 4K. An **UltraCompact** version of the model is available on the Github release. The UltraCompact version is faster, without a perceptual hit to quality in most cases. 📌 More comparisons **** **FAQ** - *How does v2 differ from v1?* I'm so glad you asked! A shortcoming of the v1 models is that the model really struggled on sources which were poorly mastered. This is unfortunately still very common even with modern anime. v2 is far more capable of dealing with such sources. - *How does Ani4K v2 differ from JaNai v3? Which one should I pick?* JaNai v3 is a fantastic model, and shares many of the fundamental training objectives behind Ani4K (DOF preservation, faithfulness, etc.). I'd say that the primary difference is one of training philosophy-- JaNai seeks to render the source as if it was originally mastered in 4K, whereas Ani4K seeks to produce an upscale that is as close as possible to the source (while cleaning up any issues). Long story short, test both models and see which you prefer. - *What model versions are there?* Ani4K comes in Compact and UltraCompact flavors. Compact is of course a standard option. UltraCompact provides a noticeable performance uplift, without too much impact on quality. I ultimately did not train a SuperUltraCompact variant as I felt the hit to the model quality was far too significant.

Simple pretrain for DAT2, trained on CC0 content. "Ethical" model

Official 4x drct-l Pretrain model, as released on their DRCT Github Page

The model was trained for some tests, but maybe someone will need to remove a screentone with a color image; in addition to the screentone, it can handle small halftones quite well

I just decided to finish off the previous one, made the manga design more clear and corrected the deviation in color temperature by adding a more diverse set of colors

Simple pretrain for RealPLKSR, trained on CC0 or self-made content. Don't use V1-V3, they're not stable for training. Dataset consisted of simple resizing: - Lanczos - Linear - Bicubic - Box

A 4x model for remove paper texture and screentone from independent manga scans. Well, I don’t even know how to describe it, a small breakthrough in my war with self-scans, I’m still not very happy with it, but maybe the lump will come in handy one day https://discord.com/channels/547949405949657098/547949806761410560/1204387665798103050

A 4x model for remove paper texture and screentone from independent manga scans. This is my first model for this, it can be said to remove poorly https://discord.com/channels/547949405949657098/1159598245417332796/1185562840304721921

## RealPLKSR GAN pretrain **Scale:** 4x **Architecture:** RealPLKSR **Download:** GDrive | Training files: GDrive **Author:** musl **License:** CC0 **Purpose:** Pretrain **Subject:** Multipurpose **Date:** 15 May 2024 **Size:** default Real-PLKSR **I/O Channels:** 3(RGB)->3(RGB) **Dataset:** Nomos-v2 **Dataset Size:** 6000 **OTF (on the fly augmentations):** No **Pretrained Model:** 4x_realplksr_mssim_pretrain **Iterations:** ~450k **Batch Size:** 2-6 **GT Size:** 128-416 **Description:** Pretrained GAN models for RealPLKSR network. Trained on downsampling-only (nearest, bilinear, bicubic, lanczos and mitchell).

Simple pretrain for SwinIR Medium, trained on CC0 content. "Ethical" model

Simple pretrain for SwinIR Small, trained on CC0 content. "Ethical" model

## RealPLKSR mssim pretrains **Scale:** 2x and 4x **Architecture:** RealPLKSR **Links:** **`4x_realplksr_mssim_pretrain.pth`** | GDrive **`2x_realplksr_mssim_pretrain.pth`** | GDrive **Author:** musl **License:** CC0 **Purpose:** Pretrain **Subject:** Multipurpose **Date:** 08 May 2024 **Size:** default Real-PLKSR **I/O Channels:** 3(RGB)->3(RGB) **Dataset:** Nomos-v2 **Dataset Size:** 6000 **OTF (on the fly augmentations):** No **Pretrained Model:** scratch **Iterations:** ~260k **Batch Size:** 2-6 **GT Size:** 128-256 **Description:** Pretrained models for RealPLKSR network. Trained on downsampling-only (nearest, bilinear, bicubic and lanczos).

This is SaurusX's original description: Fix vertical blur / split lines / shadowing. A 1x model of my 2xSwatKats. Resolves the same video problems as before, but 1x and faster and meant for chaining to other 2x models (or whatever). Input MUST be 540 vertical as the blur problem is very resolution sensitive. ------- The goal for this retrain was to get the "magic" of SwatKatsLite into SPAN for much faster video processing. It's about 80% of the way there. There's some artifacts here and there that aren't present in the original, and it has small amounts of color halos in certain areas. I've retrained this model so many times that I've just decided to release it. This particular release was trained to ~450k. The included Family Guy image shows the color ringing issue. Sorry. Update: I've added a second alt model that has less color ringing, but doesn't do quite as good with line repair. I believe I've reached architecture limits

This is SaurusX's original description: Improvement of low-quality cartoons from broadcast sources. Will greatly increase the visual quality of bad broadcast tape sources of '80s and '90s cartoons (e.g. Garfield and Friends, Heathcliff, DuckTales, etc). Directly addresses chroma blur, dot crawl, and rainbowing. You're highly advised to take care of haloing beforehand in your favorite video editor as the model will not fix it and may make existing halos more noticeable. ------- Sadly this model has some intense artifacts. Thankfully the SPAN re-train reduced these a bit, but they're still problematic. This was just a quick test of SPAN's capabilities. This was trained only to ~66k iterations, compared to the 480k of the original.

A 4x model for 4x texture image upscaler, handling jpg compression, some noise and slight blur . Link to Github Release Name: 4xTextureDAT2_otf Author: Philip Hofmann Release: 13.12.2023 License: CC BY 4.0 Network: DAT Arch Option: DAT2 Scale: 4 Iterations: 125000 batch_size: 6 HR_size: 128 Dataset: GTAV_512_Dataset Number of train images: 30122 OTF Training: Yes Pretrained_Model_G: DAT_2_x4 Description: 4x upscale texture images, trained with the Real-ESRGAN otf pipeline so handles jpg compression, some noise and slight blur

Official 4x finetune pretrain from ATD # Adaptive Token Dictionary This repository is an official implementation of the paper "Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary", CVPR, 2024. [Arxiv] [visual results] [pretrained models] By Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. > **Abstract:** Single Image Super-Resolution is a classic computer vision problem that involves estimating high-resolution (HR) images from low-resolution (LR) ones. Although deep neural networks (DNNs), especially Transformers for super-resolution, have seen significant advancements in recent years, challenges still remain, particularly in limited receptive field caused by window-based self-attention. To address these issues, we introduce a group of auxiliary **A**daptive **T**oken **D**ictionary to SR Transformer and establish an **ATD**-SR method. The introduced token dictionary could learn prior information from training data and adapt the learned prior to specific testing image through an adaptive refinement step. The refinement strategy could not only provide global information to all input tokens but also group image tokens into categories. Based on category partitions, we further propose a category-based self-attention mechanism designed to leverage distant but similar tokens for enhancing input features. The experimental results show that our method achieves the best performance on various single image super-resolution benchmarks. > > <img width="800" src="figures/tdca-acmsa.png"> > <img width="800" src="figures/arch.png"> > <br/><br/> > <img width="800" src="figures/viscomp_076.png">

4x ai generated image upscaler trained with otf The 4xLexicaDAT2_hb generated some weird lines on some edges. 4xNomosUniDAT is a different checkpoint of 4xNomosUniDAT_otf (145000), I liked the result a bit more in that example.

4x photo upscaler trained with otf

4x universal upscaler trained with otf

A 4x model.

Official 4x rgt-s pretrain from RGT # Recursive Generalization Transformer for Image Super-Resolution Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, and Xiaokang Yang, "Recursive Generalization Transformer for Image Super-Resolution", ICLR, 2024 [paper] [arXiv] [supplementary material] [visual results] [pretrained models] --- > **Abstract:** Transformer architectures have exhibited remarkable performance in image superresolution (SR). Since the quadratic computational complexity of the selfattention (SA) in Transformer, existing methods tend to adopt SA in a local region to reduce overheads. However, the local design restricts the global context exploitation, which is crucial for accurate image reconstruction. In this work, we propose the Recursive Generalization Transformer (RGT) for image SR, which can capture global spatial information and is suitable for high-resolution images. Specifically, we propose the recursive-generalization self-attention (RG-SA). It recursively aggregates input features into representative feature maps, and then utilizes cross-attention to extract global information. Meanwhile, the channel dimensions of attention matrices ($query$, $key$, and $value$) are further scaled to mitigate the redundancy in the channel domain. Furthermore, we combine the RG-SA with local self-attention to enhance the exploitation of the global context, and propose the hybrid adaptive integration (HAI) for module integration. The HAI allows the direct and effective fusion between features at different levels (local or global). Extensive experiments demonstrate that our RGT outperforms recent state-of-the-art methods quantitatively and qualitatively. ---

Model for halftone images.

Model for halftone images.

A 4x model for Pretrain to quick start new models..

Basic 1x model to use for SPAN anime models. Enjoy

Model trained only on downscale degradation (bicubic, bilinear, nearest, lanczos and mitchell). Can be used to start new Real-CUGAN models. ONNX included on the dir.

This is just a very basic 2x model to start your anime models with on SPAN, nothing special. I tried to keep it as basic as possible

Please load it using `strict_load_g: false`. Model trained only on downscale degradation (bicubic, bilinear, nearest, lanczos and mitchell). Can be used to start new SPAN models.

Model trained only on downscale degradation (bicubic, bilinear, nearest, lanczos and mitchell). Can be used to start new DCTLSA models.

Official 2x pretrain for SPAN.

Official 4x pretrain for SPAN.

Official 4x Pretrain of Dual Aggregation Transformer for Image Super-Resolution DAT Transformer has recently gained considerable popularity in low-level vision tasks, including image super-resolution (SR). These networks utilize self-attention along different dimensions, spatial or channel, and achieve impressive performance. This inspires us to combine the two dimensions in Transformer for a more powerful representation capability. Based on the above idea, we propose a novel Transformer model, Dual Aggregation Transformer (DAT), for image SR. Our DAT aggregates features across spatial and channel dimensions, in the inter-block and intra-block dual manner. Specifically, we alternately apply spatial and channel self-attention in consecutive Transformer blocks. The alternate strategy enables DAT to capture the global context and realize inter-block feature aggregation. Furthermore, we propose the adaptive interaction module (AIM) and the spatial-gate feed-forward network (SGFN) to achieve intra-block feature aggregation. AIM complements two self-attention mechanisms from corresponding dimensions. Meanwhile, SGFN introduces additional non-linear spatial information in the feed-forward network. Extensive experiments show that our DAT surpasses current methods.

Meant to quick start new DITN models trained on `neosr`. Trained only on downscale degradation (bicubic, bilinear, nearest, lanczos and mitchell). Note: Flash Attention breaks inference compatibility with official code.

Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR) While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR) While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR) While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR) While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR) While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR) While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

Official DAT2 4x Pretrain of Dual Aggregation Transformer for Image Super-Resolution DAT Transformer has recently gained considerable popularity in low-level vision tasks, including image super-resolution (SR). These networks utilize self-attention along different dimensions, spatial or channel, and achieve impressive performance. This inspires us to combine the two dimensions in Transformer for a more powerful representation capability. Based on the above idea, we propose a novel Transformer model, Dual Aggregation Transformer (DAT), for image SR. Our DAT aggregates features across spatial and channel dimensions, in the inter-block and intra-block dual manner. Specifically, we alternately apply spatial and channel self-attention in consecutive Transformer blocks. The alternate strategy enables DAT to capture the global context and realize inter-block feature aggregation. Furthermore, we propose the adaptive interaction module (AIM) and the spatial-gate feed-forward network (SGFN) to achieve intra-block feature aggregation. AIM complements two self-attention mechanisms from corresponding dimensions. Meanwhile, SGFN introduces additional non-linear spatial information in the feed-forward network. Extensive experiments show that our DAT surpasses current methods.

A 2x model.

This is my attempt at making cugan more efficient and also training it with better loss functions than official training code. Took me months to train on a 4090. It is pure pain to train this network. Nobody will probably beat my iter record any time soon. I tried to balance speed/vram and quality. The image quality is very close to cugan architecture while having ultracompact speed. It usually looks way better than compact, but has different kinds of artefacts. If you have high contrast lines which are a few or one pixel or thin lines with movement, then the model will struggle a bit. That is quite rare though. Exported as onnx for mlrt or vsgan. https://github.com/styler00dollar/VSGAN-tensorrt-docker. I also included old l1 for pretrain purposes, but be warned, it is hard to train.

A 3x model.

Official Paper model. See https://github.com/XPixelGroup/HAT for more details.

Upscale 90s and early 2000 American cartoons

This is the v2 of my 2xGT model, the main purpose of it is to improve upon the v1 model and make more it generalized. for upscaling videos that have grain I would recommend denoising and de-haloing it before passing it through the model.

Official Paper pretrain model. See https://github.com/XPixelGroup/HAT for more details.

This is my first 2xcompact model, the main purpose of it is to upscale Dragon ball GT. for upscaling videos that have grain I would recommend denoising and dehaloing it before passing it to the model for temporal stability.

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. A tiny small model for general scenes. De-noising version, meant to be interpolated with the normal one.

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. A tiny small model for general scenes.

A 4x model for Art/People. A sharp upscaler trained specifically for small sprite faces. Does not blend, so avoid using on painted/photo portraits unless you were trying to retain more of the outlines somehow.

Official paper pretrain model SRFormer: Permuted Self-Attention for Single Image Super-Resolution Yupeng Zhou 1, Zhen Li 1, Chun-Le Guo 1, Song Bai 2, Ming-Ming Cheng 1, Qibin Hou 1 1TMCC, School of Computer Science, Nankai University 2ByteDance, Singapore arXiv The official PyTorch implementation of SRFormer: Permuted Self-Attention for Single Image Super-Resolution (arxiv). SRFormer achieves state-of-the-art performance in classical image SR lightweight image SR real-world image SR The results can be found here. Abstract: In this paper, we introduce SRFormer, a simple yet effective Transformer-based model for single image super-resolution. We rethink the design of the popular shifted window self-attention, expose and analyze several characteristic issues of it, and present permuted self-attention (PSA). PSA strikes an appropriate balance between the channel and spatial information for self-attention, allowing each Transformer block to build pairwise correlations within large windows with even less computational burden. Our permuted self-attention is simple and can be easily applied to existing super-resolution networks based on Transformers. Without any bells and whistles, we show that our SRFormer achieves a 33.86dB PSNR score on the Urban100 dataset, which is 0.46dB higher than that of SwinIR but uses fewer parameters and computations. We hope our simple and effective approach can serve as a useful tool for future research in super-resolution model design. Our code is publicly available at https://github.com/HVision-NKU/SRFormer. Contents Installation & Dataset Training Testing Results Pretrain Models Citations License and Acknowledgement Installation & Dataset python 3.8 pyTorch >= 1.7.0 cd SRFormer pip install -r requirements.txt python setup.py develop Dataset We use the same training and testing sets as SwinIR, the following datasets need to be downloaded for training. Task Training Set Testing Set classical image SR DIV2K (800 training images) or DIV2K +Flickr2K (2650 images) Set5 + Set14 + BSD100 + Urban100 + Manga109 lightweight image SR DIV2K (800 training images) Set5 + Set14 + BSD100 + Urban100 + Manga109 real-world image SR DIV2K (800 training images) +Flickr2K (2650 images) + OST (10324 images for sky,water,grass,mountain,building,plant,animal) RealSRSet+5images Training Please download the dataset corresponding to the task and place them in the folder specified by the training option in folder /options/train/SRFormer Follow the instructions below to train our SRFormer. # train SRFormer for classical SR task ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_SRx2_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_SRx3_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_SRx4_scratch.yml # train SRFormer for lightweight SR task ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_light_SRx2_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_light_SRx3_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_light_SRx4_scratch.yml Testing # test SRFormer for classical SR task python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_DF2Ksrx2.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_DF2Ksrx3.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_DF2Ksrx4.yml # test SRFormer for lightweight SR task python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_light_DIV2Ksrx2.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_light_DIV2Ksrx3.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_light_DIV2Ksrx4.yml Results We provide the results on classical image SR, lightweight image SR, realworld image SR. More results can be found in the paper. The visual results of SRFormer will upload to google drive soon. Classical image SR Results of Table 4 in the paper Results of Figure 4 in the paper Lightweight image SR Results of Table 5 in the paper Results of Figure 5 in the paper Model size comparison Results of Table 1 and Table 2 in the Supplementary Material Realworld image SR Results of Figure 8 in the paper Pretrain Models Pretrain Models can be download from google drive. To reproduce the results in the article, you can download them and put them in the /PretrainModel folder. Citations You may want to cite: @article{zhou2023srformer, title={SRFormer: Permuted Self-Attention for Single Image Super-Resolution}, author={Zhou, Yupeng and Li, Zhen and Guo, Chun-Le and Bai, Song and Cheng, Ming-Ming and Hou, Qibin}, journal={arXiv preprint arXiv:2303.09735}, year={2023} } License and Acknowledgement This project is released under the Apache 2.0 license. The codes are based on BasicSR, Swin Transformer, and SwinIR. Please also follow their licenses. Thanks for their awesome works.

Official paper pretrain model SRFormer: Permuted Self-Attention for Single Image Super-Resolution Yupeng Zhou 1, Zhen Li 1, Chun-Le Guo 1, Song Bai 2, Ming-Ming Cheng 1, Qibin Hou 1 1TMCC, School of Computer Science, Nankai University 2ByteDance, Singapore arXiv The official PyTorch implementation of SRFormer: Permuted Self-Attention for Single Image Super-Resolution (arxiv). SRFormer achieves state-of-the-art performance in classical image SR lightweight image SR real-world image SR The results can be found here. Abstract: In this paper, we introduce SRFormer, a simple yet effective Transformer-based model for single image super-resolution. We rethink the design of the popular shifted window self-attention, expose and analyze several characteristic issues of it, and present permuted self-attention (PSA). PSA strikes an appropriate balance between the channel and spatial information for self-attention, allowing each Transformer block to build pairwise correlations within large windows with even less computational burden. Our permuted self-attention is simple and can be easily applied to existing super-resolution networks based on Transformers. Without any bells and whistles, we show that our SRFormer achieves a 33.86dB PSNR score on the Urban100 dataset, which is 0.46dB higher than that of SwinIR but uses fewer parameters and computations. We hope our simple and effective approach can serve as a useful tool for future research in super-resolution model design. Our code is publicly available at https://github.com/HVision-NKU/SRFormer. Contents Installation & Dataset Training Testing Results Pretrain Models Citations License and Acknowledgement Installation & Dataset python 3.8 pyTorch >= 1.7.0 cd SRFormer pip install -r requirements.txt python setup.py develop Dataset We use the same training and testing sets as SwinIR, the following datasets need to be downloaded for training. Task Training Set Testing Set classical image SR DIV2K (800 training images) or DIV2K +Flickr2K (2650 images) Set5 + Set14 + BSD100 + Urban100 + Manga109 lightweight image SR DIV2K (800 training images) Set5 + Set14 + BSD100 + Urban100 + Manga109 real-world image SR DIV2K (800 training images) +Flickr2K (2650 images) + OST (10324 images for sky,water,grass,mountain,building,plant,animal) RealSRSet+5images Training Please download the dataset corresponding to the task and place them in the folder specified by the training option in folder /options/train/SRFormer Follow the instructions below to train our SRFormer. # train SRFormer for classical SR task ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_SRx2_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_SRx3_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_SRx4_scratch.yml # train SRFormer for lightweight SR task ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_light_SRx2_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_light_SRx3_scratch.yml ./scripts/dist_train.sh 4 options/train/SRFormer/train_SRFormer_light_SRx4_scratch.yml Testing # test SRFormer for classical SR task python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_DF2Ksrx2.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_DF2Ksrx3.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_DF2Ksrx4.yml # test SRFormer for lightweight SR task python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_light_DIV2Ksrx2.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_light_DIV2Ksrx3.yml python basicsr/test.py -opt options/test/SRFormer/test_SRFormer_light_DIV2Ksrx4.yml Results We provide the results on classical image SR, lightweight image SR, realworld image SR. More results can be found in the paper. The visual results of SRFormer will upload to google drive soon. Classical image SR Results of Table 4 in the paper Results of Figure 4 in the paper Lightweight image SR Results of Table 5 in the paper Results of Figure 5 in the paper Model size comparison Results of Table 1 and Table 2 in the Supplementary Material Realworld image SR Results of Figure 8 in the paper Pretrain Models Pretrain Models can be download from google drive. To reproduce the results in the article, you can download them and put them in the /PretrainModel folder. Citations You may want to cite: @article{zhou2023srformer, title={SRFormer: Permuted Self-Attention for Single Image Super-Resolution}, author={Zhou, Yupeng and Li, Zhen and Guo, Chun-Le and Bai, Song and Cheng, Ming-Ming and Hou, Qibin}, journal={arXiv preprint arXiv:2303.09735}, year={2023} } License and Acknowledgement This project is released under the Apache 2.0 license. The codes are based on BasicSR, Swin Transformer, and SwinIR. Please also follow their licenses. Thanks for their awesome works.

A 1x model for Skin Upscaling. Adds plausible high frequency detail and removes subtle blur. This is an early unfinished attempt at a x1 Lite model designed specifically for enhancing detail on skin diffuse textures of 3d characters. Even in its current state it works quite well. Best suited for uncompressed or cleaned textures - otherwise it may just enhance any existing compression artefacts too. The training set included faces, body parts, eyes and hair in a variety of skin types and tones so it should work well on most related diffuse textures. However it's not just limited to skin, many other images and textures can be enhanced with this model. The results are subtle, so run multiple times if desired. Pretrained model: 50/50 Interpolation of DIV2K-Lite and SpongeBC1-Lite

A 1x model for Correct old color photos that are tinted red. Correct old color photos that are tinted red. Model only tested in nature.

Morrowind Mod Textures

A 2x model for Animethemes. I think one of sharpest models for anime. Sample video: https://cdn.discordapp.com/attachments/579685650824036387/1002663424695750686/output.mp4 Plz don't steal without credits, k thx.

A 4x model for Anime PSP games, Fate Extra. This model was trained as a favor to Demon and the Fate Extra community. It leaves a nice grain on the images and upscales lines and details accurately without looking odd. This model works on most anime-style PSP games. Enjoy! It works best on content with dithering and quantization. NOTICE: I have included both NCNN and ONNX models to make upscaling easier if you rely on either of these. For NCNN, there's two versions. One is FP16 and the other is FP32. FP16 works best on RTX GPUs. Choose FP32 if in doubt about compatibility, or if FP16 doesn't work for you. To use ONNX, download chaiNNer and upscale through there with the ONNX nodes.

A 2x model. Pretrained using Pretrained_Model_G: RealESRGAN_x4plus_anime_6B.pth /. RealESRGAN_x4plus_anime_6B.pth (sudo_RealESRGAN2x_3.332.758_G.pth) Tried to make the best 2x model there is for drawings. I think i archived that. And yes, it is nearly 3.8 million iterations (probably a record nobody will beat here), took me nearly half a year to train. It can happen that in one edge is a noisy pattern in edges. You can use padding/crop for that. I aimed for perceptual quality without zooming in like 400%. Since RealESRGAN is 4x, I downscaled these images with bicubic. I would recommend my VSGAN code though and just load the onnx. https://github.com/styler00dollar/VSGAN-tensorrt-docker I just wanted a good 2x model for animations, but that model can also be used for wallpapers and so on. Before I hear people complaining, the dropout model is a modified architecture. Stuff like cupscale or chaiNNer won't work with pth. Load the onnx with VSGAN or chaiNNer. I did add the model before switching to dropout though, which is normal ESRGAN pth, that one should work everywhere. I also converted everything into onnx, jit and ncnn, so pretty much everything there is. If you want to use ncnn, don't use nihuis code (that also includes cupscale), these codes don't include propper tiling in C++, which is very bad for this model. I think chaiNNer should have overlap/padding with ncnn, so use that instead if you really want ncnn. Plz don't steal without credits, k thx.

Tried to make the best 2x model there is for drawings. I think i archived that. And yes, it is nearly 3.8 million iterations (probably a record nobody will beat here), took me nearly half a year to train. It can happen that in one edge is a noisy pattern in edges. You can use padding/crop for that. I aimed for perceptual quality without zooming in like 400%. Since RealESRGAN is 4x, I downscaled these images with bicubic. I would recommend my VSGAN code though and just load the onnx. https://github.com/styler00dollar/VSGAN-tensorrt-docker I just wanted a good 2x model for animations, but that model can also be used for wallpapers and so on. Before I hear people complaining, the dropout model is a modified architecture. Stuff like cupscale or chaiNNer won't work with pth. Load the onnx with VSGAN or chaiNNer. I did add the model before switching to dropout though, which is normal ESRGAN pth, that one should work everywhere. I also converted everything into onnx, jit and ncnn, so pretty much everything there is. If you want to use ncnn, don't use nihuis code (that also includes cupscale), these codes don't include propper tiling in C++, which is very bad for this model. I think chaiNNer should have overlap/padding with ncnn, so use that instead if you really want ncnn. Plz don't steal without credits, k thx.

A 2x model for Realtime animation restauration and doing stuff like deblur and compression artefact removal. (Teacher: RealESRGANv2-animevideo-xsx2.pth) My first attempt to make a REALTIME 2x upscaling model while also applying teacher student learning. It beats Anime4k in every way. These benchmarks use a 3060ti and it shows that everything better than a 3060ti should be able to handle 1080p input if you create engine files and use my TensorRT code. You can see in the readme how to convert onnx files into engines. The 2 right bars compare normal Compact2 and Ultracompact in speed, the 2 on the left showcase older apis I used which isn't too important for this showcase. To use this, you need to use my code which is https://github.com/styler00dollar/VSGAN-tensorrt-docker. If you use Manjaro, it is also possible to pipe the data stream directly into mpv, so you can watch it in a video player without rendering a video. Yeah the model does seem a little noisy if you zoom in a lot, but don't forget that the model itself is only 1.2mb. I think it does quite well. I still try to improve on fast models, but this is good enough to share as a first model. Plz don't steal without credits, k thx.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

This is a collection of pretrained models for Real-ESRGAN's Compact architecture. There are 1x, 2x, and 4x models, as well as 1x and 2x "UltraCompact" and "SuperUltraCompact" models (think of these as the equivalent to ESRGAN "lite" models). By using these are pretrains for your models, you can ensure that your models are able to be interpolated with other Compact models that were trained from these. These pretrains are compatible with most existing compact models.

A 2x model for Animation interpolation. I never really mentioned it in model-releases prior since I think not too many care about interpolation here, but I trained a rife4 model for animation a some months ago, which is better than rife4 and rife4.2 imo. Thought I should also mention it here as well. I also converted it into ncnn. (Nihuis rife ncnn models are only exported with the fastest mode and not the best quality. I exported ncnn models for the most important quality settings. Due to different export/quality settings, there are multiple models. For that reason alone, my ncnn models are much better too, since nihui only exported the fastest one.) My https://github.com/styler00dollar/VSGAN-tensorrt-docker also has the rife ncnn extention, which can use VMAF, dedup, scene detection and so on, which I would recommend. My models are in that extention as well, just select model 10, 11 or 12 and use the dev docker. That test video is done with 2x framerate, enbemble True and FastMode False, combined with scene detection and dedup stuff, tta False. Towards the best quality rife can do. Plz don't steal without credits, k thx. Pretrained model: RifeV4 Sample video: https://cdn.discordapp.com/attachments/579685650824036387/990345004260151296/ngnl_sudorife.mp4

This model is intended to mainly handle PS2 compression and a mixture of Realistic and cartoonish textures, it's not meant to be used for very low resolution textures such as item icons.

A 1x model for Anti aliasing / Deblur. To eliminate aliasing and general artifacts caused by not enough resolution while bringing out details Im stopping its training here because it's getting worse, i think of some aligment issues by game's rendering pipeline + downscaling...

A 1x model for Anti aliasing / Deblur. To eliminate aliasing and general artifacts caused by not enough resolution while bringing out details Im stopping its training here because it's getting worse, i think of some aligment issues by game's rendering pipeline + downscaling...

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. We add small models that are optimized for anime videos :-)

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. This model is optimized for anime images with much smaller model size.

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data. We add small models that are optimized for anime videos :-)

This model was trained on lossless video frames of Metal Arms: Glitch in the System, compressed with Bink V1 for the LR frames. It's a lot more efficient than my first DeBink model, and also has less artifacts. It's not quite as robust, but the compression is barely noticeable in videos after processing with this. Sample video: https://cdn.discordapp.com/attachments/903415274521374750/904129653374078976/XB_Demo_Mov-85600_G.mp4

Attempts to remove the damages done from noise. Successor of sorts to Noisetoner_Poisson_150000_G

These models deblur most images. It was trained mostly on aniso2 and iso blurring (BSRGAN augmentation) with some gaussian mixed in. It performs well on most blurry images, but I'd recommend using something like Fatality_Deblur for very strong gaussian blur.

Remove film grain/noise

Swift-SRGAN - Rethinking Super-Resolution for real-time inference In recent years, there have been several advancements in the task of image super-resolution using the state of the art Deep Learning-based architectures. Many super-resolution-based techniques previously published, require high-end and top-of-the-line Graphics Processing Unit (GPUs) to perform image super-resolution. With the increasing advancements in Deep Learning approaches, neural networks have become more and more compute hungry. We took a step back and, focused on creating a real-time efficient solution. We present an architecture that is faster and smaller in terms of its memory footprint. The proposed architecture uses Depth-wise Separable Convolutions to extract features and, it performs on-par with other super-resolution GANs (Generative Adversarial Networks) while maintaining real-time inference and a low memory footprint. A real-time super-resolution enables streaming high resolution media content even under poor bandwidth conditions. While maintaining an efficient trade-off between the accuracy and latency, we are able to produce a comparable performance model which is one-eighth (1/8) the size of super-resolution GANs and computes 74 times faster than super-resolution GANs. NOTE: The author used the wrong file extensions for the models *on GitHub*. You will download a `.pth.tar` file. This is not actually a TAR file. Change the file extension to just `.pth` and the model will work.

Swift-SRGAN - Rethinking Super-Resolution for real-time inference In recent years, there have been several advancements in the task of image super-resolution using the state of the art Deep Learning-based architectures. Many super-resolution-based techniques previously published, require high-end and top-of-the-line Graphics Processing Unit (GPUs) to perform image super-resolution. With the increasing advancements in Deep Learning approaches, neural networks have become more and more compute hungry. We took a step back and, focused on creating a real-time efficient solution. We present an architecture that is faster and smaller in terms of its memory footprint. The proposed architecture uses Depth-wise Separable Convolutions to extract features and, it performs on-par with other super-resolution GANs (Generative Adversarial Networks) while maintaining real-time inference and a low memory footprint. A real-time super-resolution enables streaming high resolution media content even under poor bandwidth conditions. While maintaining an efficient trade-off between the accuracy and latency, we are able to produce a comparable performance model which is one-eighth (1/8) the size of super-resolution GANs and computes 74 times faster than super-resolution GANs. NOTE: The author used the wrong file extensions for the models *on GitHub*. You will download a `.pth.tar` file. This is not actually a TAR file. Change the file extension to just `.pth` and the model will work.

A 2x model for Anime - interpolation.

A 4x model for Game textures/Art. This model upscales artwork for the game Volfoss (2001). The model peaked in quality very quickly. The NR model removes most noise, but has the downside of removing transparent portions. Use the main model in most cases.

Meant as an experiment to test latest techniques implemented on traiNNer, including: AdaTarget, KernelGAN, UNet discriminator, nESRGAN+ arch, noise patches, camera noise, isotropic/anisotropic/sinc blur, frequency separation, contextual loss, mixup, clipL1 pixel loss, AdamP optimizer, etc. The config file is provided on the download link above. I encourage everybody to mirror the model, distribute and modify it in anywway you want.

A 2x model for Animation, Pixel Art. This model has VERY strong compression removal and line restructuring that allows it to restore any heavily compressed drawings, animation, cartoons, or anime. Also works on games as well as DDS compression. It renders frames very quickly and is very viable for restoring videos. Pretrained model: 4x-MMScale (Yes, 4x. mistakes were made)

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

SwinIR: Image Restoration Using Swin Transformer Image restoration is a long-standing low-level vision problem that aims to restore high-quality images from low-quality images (e.g., downscaled, noisy and compressed images). While state-of-the-art image restoration methods are based on convolutional neural networks, few attempts have been made with Transformers which show impressive performance on high-level vision tasks. In this paper, we propose a strong baseline model SwinIR for image restoration based on the Swin Transformer. SwinIR consists of three parts: shallow feature extraction, deep feature extraction and high-quality image reconstruction. In particular, the deep feature extraction module is composed of several residual Swin Transformer blocks (RSTB), each of which has several Swin Transformer layers together with a residual connection. We conduct experiments on three representative tasks: image super-resolution (including classical, lightweight and real-world image super-resolution), image denoising (including grayscale and color image denoising) and JPEG compression artifact reduction. Experimental results demonstrate that SwinIR outperforms state-of-the-art methods on different tasks by up to 0.14~0.45dB, while the total number of parameters can be reduced by up to 67%.

A 2x model for Anime - interpolation. Sample video: https://cdn.discordapp.com/attachments/724353640521007145/873861876352704533/1.webm

A 1x model. Pretrained using 4xPSNR. Texture retaining denoiser and sharpener for digital artwork Sample images: https://1drv.ms/u/s!Aip-EMByJHY28xKFp93VIGydNsoN?e=vN2WYh

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data.

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data.

Upscaling of anime art (Specifically visual novel CG art)

A 2x model for RL videos - interpolation. Architecture: CAIN (CAIN 1 Group) Trained mostly on abba music videos

A 4x model for Upscales ground textures.

Remove JPEG artifacts from manga without destroying screentone and other details. For 80-50 JPEG Quality

Remove JPEG artifacts from manga without destroying screentone and other details. For 95-80 JPEG Quality

Remove JPEG artifacts from manga without destroying screentone and other details. For 60-30 JPEG Quality

Basic pixel art upscaling, for people who want a more simpler style and lightweight pixel art upscaling model. Sample images: https://drive.google.com/drive/folders/1mYTMpwDlKQulBmjgKpcxL3-T6xAGXVxl?usp=sharing

This model removes early 2000s Bink and other compression artifacts. Works well on almost any image or video compression type.

This model removes early 2000s Bink and other compression artifacts. Works well on almost any image or video compression type.

This model removes early 2000s Bink and other compression artifacts. Works well on almost any image or video compression type.

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

Upscale Anime while keeping as much of the original detail as possible. Isn't among the sharpest models and if the anime is somewhat blurry it will retain that detail but add more pixels to give a smoother edge. Doesn't change the color very much and may make some scene's very slightly darker. Pretrained model: Pony Models

A 2x model for Video Frame Interpolation for anime openings. Pretrained using My ~~first~~ (ok technically second) attempt to create a video frame interpolation model and I like how it turned out. To use it, you can either use https://gitlab.com/hubert.sontowski2007/cainapp, https://github.com/styler00dollar/Colab-CAIN or the bot in [Game Upscale] (model called rvpv1 there, just use --model rvpv1). Some demo videos in pcloud, but you need to download them. The web player seems to playback in low fps. The architecture is mainly the same to the original CAIN, but i modified the padding to be zero padding instead. ``.pt`` means JIT model, `.pth`, means normal pytorch model. Architecture file is in pcloud as well. And no, no cupscale or flowframes.. Sample video: https://files.catbox.moe/xiq9vi.mp4

A 4x model for Art. Illustrations with with larger shaped features (?). Pretrained model: 4x_UltraFArt

A 4x model for Art. Illustrations with with larger shaped features (?). Pretrained model: 4x_UltraFArt

A 4x model for Art. Illustrations with with larger shaped features (?). Pretrained model: 4x_UltraFArt

A 4x model for Art. Illustrations with with larger shaped features (?). Pretrained model: 4x_UltraFArt

Sharpening, line darkening and slight line thinning, specifically made for the Boruto anime.

Upscale MLP episodes. Full version. Able to handle compression better compared to V1.0 and no longer creates halo's/rainbows. Good for Vector 2D art however converts detail to blobs.

Upscale MLP episodes. Full version. Able to handle compression better compared to V1.0 and no longer creates halo's/rainbows. Good for Vector 2D art however converts detail to blobs.

Upscale MLP episodes. Full version. Able to handle compression better compared to V1.0 and no longer creates halo's/rainbows. Good for Vector 2D art however converts detail to blobs.