BCGone Smooth

To retain details while removing artifacts caused by dxt compression on textures

DEDXT

Removing compression artifacts from DXT1 compressed textures. This model was created to remove DXT1 compression artifacts from textures imported into the Source Engine. Compressed textures in the engine sometimes have a green-tint which this model also corrects. The data for this model contained a good mix of diffuse textures and normal maps which means this model is pretty good at removing compression from normals as well. Creating this model was a real learning experience for me and I hope someone finds a good use for it.

DXTDecompressor Source V3

This model is made for Source Engine textures. It tries to remove compression artifacts such as blockyness, discoloration, green tint. It does pretty well on a lot of things, realistic stuff as well, but it was mostly made to work on TF2 textures. It's made to keep as much detail as possible, without any unnecessary denoising/sharpening. Huge thanks to Twittman for assisting me along this journey.

DXTless SourceEngine

Gamecube and Wii textures Dumps.

It is designed as a faster alternative to the full size model, it requires much less Vram and is significantly faster.

HDCube Compact

Category: Fabric/Cloth
Purpose: Fabric

This model set upscales fabric or cloth textures (works on cats too!). The Alt model is just an earlier iteration version. It may work better on some images.The images need to be minimally compressed or passed through a decompression model first. It works with DDS compression though.

Fabric

Fabric-Alt

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.

GameAI 2.0

Purpose: Gamecube and Wii textures.

Gamecube and Wii textures (mainly DXT and 8bit color compression). Is good for preserving fine details without affecting the original style too much, it is not suitable for pixel art, small icons and text under 16 pixel.

Pretrained model: 4x_NMKD Siax

HDCube

Gamecube and Wii textures Dumps.

https://github.com/Venomalia/HDcube/tree/main/v2

HDCube2

Purpose: Gamecube and Wii textures.
Pretrained: 4x_HDcube2

It can be used for all image formats supported by Gamecube and Wii hardware and can remove its typical artifacts like CMPR Block Compression (DXT1 algorithm, also known as BC1), color palette errors, color reduction up to 8bit color depth and 1bit alpha depth.

HDCube3

4x anime image upscaler with a bit of otf jpg compression and blur. Trained on musl's hfa2k dataset release for Anime SISR, which has been extracted from modern anime films, where the selection criteria was high SNR, no DOF and high frequency information.

Examples: https://imgsli.com/MTc2NDgx (PS Example input images are small, each around 500x280 px) \
Example input files: https://drive.google.com/drive/folders/1RI6gGqRy-KxDujbaIrpEMdvFkIWHvfPt \
Example output files: https://drive.google.com/drive/folders/1GqHwPlFp6bIQl4R1AxmrmJ6vUUD8FUxH \
All my model files can be found on my github repo https://github.com/Phhofm/models

HFA2k

Purpose: 4x anime upscaler handling video compression artifacts, trained for 200 epochs

4x anime upscaler handling video compression artifacts since trained with otf degradations for "mpeg2video", "libxvid", "libx264", "libx265" with crf 20-32, mpeg bitrate 3800-5800 (together with the standard Real-ESRGAN otf pipeline). A faster arch using this otf degradation pipeline would be great for handling video compression artifacts. Since this one is a GRL model and therefore slow, [as noted by the dev](https://github.com/Kiteretsu77/VCISR-official/issues/3) maybe more for research purposes (or more for single images/screenshots). Trained using [VCISR](https://github.com/Kiteretsu77/VCISR-official) for 200 epochs.  

"This is epoch 200 and the start iteration is 85959 with learning rate 2.5e-05"  

Slow Pics examples:  
[h264_crf28](https://slow.pics/c/fqOnxXep)  
[ludvae1](https://slow.pics/c/OcukRy7d)  
[ludvae2](https://slow.pics/c/FI33mX4F)

HFA2k VCISR GRLGAN ep200

[Link to Github Release](https://github.com/Phhofm/models/releases/tag/4xHFA2k_ludvae_realplksr_dysample)

4xHFA2k_ludvae_realplksr_dysample  
Scale: 4  
Architecture: [RealPLKSR with Dysample](https://github.com/muslll/neosr/?tab=readme-ov-file#supported-archs)  
Architecture Option: [realplksr](https://github.com/muslll/neosr/blob/master/neosr/archs/realplksr_arch.py)  

Author: Philip Hofmann  
License: CC-BY-0.4 
Purpose: Restoration  
Subject: Anime  
Input Type: Images  
Release Date: 13.07.2024  

Dataset: HFA2k_LUDVAE  
Dataset Size: 10'272  
OTF (on the fly augmentations): No  
Pretrained Model: [4xNomos2_realplksr_dysample](https://github.com/Phhofm/models/releases/tag/4xNomos2_realplksr_dysample)  
Iterations: 165'000  
Batch Size: 12  
GT Size: 256  

Description:  
A Dysample RealPLKSR 4x upscaling model for anime single-image resolution.  
The dataset has been degraded using DM600_LUDVAE, for more realistic noise/compression. Downscaling algorithms used were imagemagick box, triangle, catrom, lanczos and mitchell. Blurs applied were gaussian, box and lens blur (using chaiNNer). Some images were further compressed using -quality 75-92. Down-up was applied to roughly 10% of the dataset (5 to 15% variation in size). Degradations orders were shuffled, to give as many variations as possible.  

Examples are inferenced with [neosr](https://github.com/muslll/neosr) testscript and the released pth file. I include the test images also as a zip file in this release together with the model outputs, so others can test their models against these test images aswell to compare.  

onnx conversions are  static since dysample doesnt allow dynamic conversion, I tested the conversions with [chaiNNer](https://github.com/chaiNNer-org/chaiNNer).  

Showcase:  
[Slowpics](https://slow.pics/c/FKDZAcyI)

4xHFA2k_ludvae_realplksr_dysample

Photos/realistic 3D with JPEG compression, quality 85-95 and 4:2:0 chroma subsampling. - Created for Myst3 images, since all have 4:2:0 chroma subsampling and existing JPEG models did not give good results. Favors smoothing over over-sharpening.

Jaypeg90

Upscale small photos with compression to 4x their size.

Trying to extend my previous model to be able to handle compression (JPG 100-30) by manually altering the training dataset, since 4xLSDIRCompact can't handle compression. Use this instead of 4xLSDIRCompact if your photo has compression (like an image from the web).

LSDIR Compact C

A 4x upscaler for BC1-compressed normal maps with a zeroed-out B channel. The output normals will also have a constant-zero B channel. Use external software or image editing plugins to properly normalize the generated normals and generate the Z component (if necessary). E.g. chaiNNer can do this with Normalize Normal Map node. Do not rely on this network producing unit vectors. 

**Dataset transformation:** \
The LR have been compressed with various BC1 compression settings (dithering, weighting) using Texconv version 2021.11.8.1. Since the contents of the B channel influence the R and G channels during compression, the LR were given a B channel that was either the Z component of the normal, a random constant color, or a random texture (usually the G channel of the associated albedo) before compression. The LR + B channel was then compressed and the resulting BC1-compressed DDS was converted back into a PNG and had its B channel zeroed out.

More information: https://github.com/RunDevelopment/ESRGAN-models/blob/main/normals/README.md

Normal RG0 BC1

A 4x upscaler for BC7-compressed normal maps with a zeroed-out B channel and no alpha. The input is required to have no alpha and constant-zero blue channel. BC7-compressed images with an alpha channel should be handled by the BC1 model instead. This is because BC7 achieves very different quality for images with and without an alpha channel. The output normals will also have a constant-zero B channel. The output normals will also have a constant-zero B channel. Use external software or image editing plugins to properly normalize the generated normals and generate the Z component (if necessary). E.g. chaiNNer can do this with Normalize Normal Map node. Do not rely on this network producing unit vectors. 

**Dataset transformation:** \
The LR have been compressed using Texconv version 2021.11.8.1. Since the contents of the B channel influence the R and G channels during compression, the LR were given a B channel that was either the Z component of the normal, a random constant color, or a random texture (usually the G channel of the associated albedo) before compression. The LR + B channel was then compressed and the resulting BC7-compressed DDS was converted back into a PNG and had its B channel zeroed out.

More information: https://github.com/RunDevelopment/ESRGAN-models/blob/main/normals/README.md

Normal RG0 BC7

PBRify Github: https://github.com/Kim2091/PBRify_Remix

Release Link: https://github.com/Kim2091/Kim2091-Models/releases/tag/4x-PBRify_RPLKSRd_V3

This update brings a new upscaling model, 4x-PBRify_RPLKSRd_V3. This model is roughly 8x faster than the current DAT2 model, while being *higher quality*. It produces far more natural detail, resolves lines and edges more smoothly, and cleans up compression artifacts better.

As a result of those improvements, PBR is also much improved. It tends to be clearer with less defined artifacts. 

However, this model is currently **only compatible with ComfyUI**. chaiNNer has not yet been updated to support this architecture.

[More Comparisons](https://imgsli.com/Mjk5NjQ5)

4x-PBRify_RPLKSRd_V3

Project page: https://github.com/Kim2091/PBRify_Remix

Yet another model in the PBRify_Remix series. This is a new upscaler to replace the previous 4x-PBRify_UpscalerSIR-M_V2 model.

This model far exceeds the quality of the previous, with far more natural detail generation and better reconstruction of lines and edges.

Comparison: https://slow.pics/c/DCjlXPGb

4x-PBRify_UpscalerDAT2_V1

This is part of a larger model set, PBRify_Remix. PBRify_Remix is an easy way to upscale and generate PBR textures using existing low quality textures. The dataset consists of ethically sourced textures from websites like ambientCG and Polyhaven, which sets it apart from most other models.

It's intended for use with RTX Remix (hence the name) but it'll work for other things as well.

4x-PBRify_UpscalerSPANV4

This model was made to upscale realistic low-res textures that are compressed by either JPEG or BC1. From my testing, this works rather well on realistic GameCube textures such as the ones from Shrek Extra Large and the board textures from Mario Party 4. This model could also work on some real life images, especially the ones that are taken outdoors.

RealisticRescaler

Pretrained: 4xESRGAN

This is my best model yet! It generates lots and lots of detail and leaves a nice texture on images. It works on most images, whether compressed or not. It does work best on JPEG compression though, as that's mostly what it was trained on. It has the ability to restore highly compressed images as well! If you want a more balanced output, check out the UltraMix Collection down below. It's a bunch of interpolated models based around UltraSharp and my other models

UltraSharp

UniScale_Restore has strong compression removal that helps with restoring heavily compressed or noisy images. It is intended to compete with BSRGAN. Trained with BSRGAN_Resize and Combo_Noise in traiNNer.

UniScale Restore

A 4x upscaler for BC1-compressed wood textures. The model was trained on wood planks, wood stems, and a bit of tree bark. It performs reasonably well on woody textures and produces sharp outputs. However, it also tends to slightly shift the hue of the image, so results might appear slightly more red.

More information: https://github.com/RunDevelopment/ESRGAN-models/blob/main/wood/README.md

Architecture	ESRGAN
Scale	4x
Size	64nf23nb
Color Mode	RGB
License	CC-BY-NC-SA-4.0 Private use Distribution Modifications Credit required Same License State Changes No Liability & Warranty Disclaimer
Date	2021-09-15
Dataset	Custom Dataset consisting of edited RAW images taken by myself. Macro shots of fabrics and various other materials.
Dataset size	346
Training iterations	53600
Training epochs	3350
Training batch size	4
Training HR size	64

Fabric-Alt

Similar Models