Skip to content
This repository has been archived by the owner on Sep 21, 2022. It is now read-only.

Model Database

victorca25 edited this page Aug 13, 2019 · 60 revisions

Outdated ESRGAN Models

The wiki has been moved, here is the new Model Database.

Upscaling - Drawings

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Manga109Attempt kingdomakrillic 4 Anime / Manga ? 4 ? ? 0.1K Manga109 RRDB_PSNR_x4
Falcon Fanart LyonHrt 4 Anime / Manga 125K 8 128 ? 3.393K Falcon Fanart RRDB_PSNR_x4
Comic Book LyonHrt 4 Comic / Drawings 115K 8 128 592 1.548K Custom (Spider-Man) none
ad_test_tf PRAGMA 4 Cartoon / Netflix 5K 16 128 ? 30K Custom (American Dad) PSNRx4
De-Toon LyonHrt 4 Toon Shading / Sprite 225K 8 128 525 7.117K Custom Cartoon-style photos RRDB_PSNR_x4
Unholy02 DinJerr 4 Anime / Manga ? ? ? ? ? CG-Painted Anime Several, see notes
Unholy03 DinJerr 4 Anime / Manga ? ? ? ? ? CG-Painted Anime Several, see notes
WaifuGAN v3 DinJerr 4 Anime / Manga 30K 2 128 ? 0.173K CG-Painted Anime Manga109v2
Lady0101 DinJerr 4 Anime / Manga 208K ? ? ? ~7K CG-Painted Anime WaifuGAN v3
DigitalFrames Klexos 4 Digital Cartoon 1.06M 15 128 96.275 0.25K - 2.5K Digital Cartoon Images RRDB_PSNR_x4

Manga109Attempt is slightly blurry, but performs well as a general upscaler.

Falcon Fanart tries to improve upon it with the goal of removing checkerboard patterns / and dithering. It has oil colour based shading with sharp lines.

The Comic Book model was trained using stills from the film spiderman into the spiderverse, has a comic book crosshatch shading effect to the images. Sample

The ad_test_tf model was designed for upscaling American Dad NTSC DVD frames (originally at 480p) to match the quality and style of Netflix's equivalent 1080p WEB-DL, which includes a slight desaturation of colors.

De-Toon, is a model that does the opposite of tooning an image. It takes toon style shading and detail, and attempts to make it realistic. Its very sensitive, and can be used on small sprites, to large images. Also included is a alt version, which is less sharp.

Unholy02 and Unholy03 were created by interpolating a whole bunch of models about 30 times, mainly with the Dinjerr's own WaifuGAN model and RRDB_esrgan. It's intended for upscaling CG-painted anime images with light outlines and produces sharper, cleaner, and more aggressive results than manga109, but may produce unnecessary outlines or details when faced with noise, so be wary of jpegs.

WaifuGAN v3 is Dinjerr's third attempt at training from a mostly anime dataset sourced from image boards and is intended for upscaling CG-painted anime with variable outlines. Only PNGs were used, mainly with brush strokes and gradients. Texturised images avoided as much as possible. If too generative, tone down by interpolating with a softer model.

Lady0101 was trained on digital paintings of ladies (mostly). Strong anti-staircasing, mediocre undithering and slight blending. It is meant to be used to upscale pixel art/paintings and transform it into digital painting style.

Upscaling - Realistic (photos, prerendered 3D, etc)

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Box buildist 4 GNU GPLv3 Realstic 390K 8 192 268 11.577K Flickr2K+Div2K+OST PSNR model from same data
Ground ZaphodBreeblebox 4 Ground Textures 305K ? 128 ? ? Custom (Ground textures Google) ?
Misc alsa64 4 GNU GPLv3 Surface Textures 220K 32 128 338 20.797K Custom (Photos) Manga109Attempt

Box was meant to be an improvement on the RRDB_ESRGAN_x4 model (comparison). It's also trained on photos, but with a much larger dataset which was downscaled with linear interpolation (box filter) instead of bicubic.

The Ground model was trained on various pictures of stones, dirt and grass using Google's image search.

The Misc model is trained on various pictures shoot by myself, including bricks, stone, dirt, grass, plants, wook, bark, metal and a few others.

Upscaling - Characters and Faces

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Trixie LyonHrt 4 Star Wars 275K 8 192 87 19.814K ? None
Face Focus LyonHrt 4 Face De-blur 275K 8 192 455 4.157K Custom (Faces) RRDB_PSNR_x4
Face Twittman 4 Face Upscaling 250K 10 128 967 3.765K Custom (Faces) 4xESRGAN

Trixie was made to bring balance to the force... Also to upscale character textures for star wars games, including the heroes, rebels, sith and imperial. Plus a few main aliens...Why called trixie? Because jar jars big adventure would be too long of a name...This also provides good upscale for face textures for general purpose as well as basic star wars textures.

The Face Focus modes was designed for slightly out of focus / blurred images of faces. It is aimed at faces / hair, but it can help to improve other out of focused images too as always just try it.

Upscaling - Pixel Art

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Xbrz LyonHrt 4 Xbrz style pixel art upscaler 90K 8 128 368 1.897K custom xbrz up-scaled RRDB_PSNR_x4
Xbrz+DD LyonHrt 4 Xbrz style pixel art upscaler with de-dithering 90K 8 128 470 1.523K custom de-dithered xbrz xbrz
ScaleNX LyonHrt 4 Scalenx style pixel art upscaler 80K 8 128 599 1.070K custom scalenx up-scaled from retroarch shader RRDB_PSNR_x4
Fatality twittman 4 (dithered) spirites 265K 10 128 160 19.7K ? Face
Rebout LyonHrt 4 Character Sprites 325K 8 128 106 23.808K Custom prepared sprites from kof 94 rebout Detoon

Fatality is meant to be used for upscaling medium resolution Sprites, dithered or undithered, it can also upscale manga/anime and gameboy camera images.

Rebout is trained to give detail to character models, with faces and hands improved. Based on the snk game kof94 rebout, although best for snk style games, does work on a variety of sprites. Also included is a interpolated version that may provide a cleaner upscale for certain sprites.

Upscaling - Specialized

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Map LyonHrt 4 Map / Old Paper with text 120K 8 192 361 2.311K Custom(Scans) none
Forest LyonHrt 4 Wood / Leaves 160K 8 192 590 2.2K Custom(?) none
Skyrim Armory alsa64 4 GNU GPLv3 Armor, Clothes and Weapons 80K 26 128 2.6K 0.8K Skyrim Mod textures Manga109Attempt
Skyrim Wood Laeris 4 Wood 75K ? ? ? ? ? ?
Skyrim Misc Deorder 4 Skyrim Diffuse Textures 105K ? 128 ? ? Skyrim Diffuse Textures ?
Fallout 4 Weapons Bob 4 Fallout Weapon Diffuse Textures 120K 13 128 2.973K 532(OTF) Fallout 4 HDDLC Weapon Diffuse Textures Manga109Attempt

The map model was trained on maps, old documents, papers and various styles of typefaces/fonts. Based on a dataset contributed by alsa64. Sample

The Forest model is focused on trees, leaves, bark and stone can be used for double upscaling for even more detail. Sample

The Armory model was trained with modded textures form Skyrim, including Clothing, Armor and Weapons. (Leather, Canvas and Metal should all work - maybe too sharp so interpolate)

The wood model was trained for Skyrim by Laeris.

The Skyrim Diffuse models is supposed to be used with Skyrim's diffuse textures. It is a bit too sharp so I recommend to interpolating with the RDDB_ESRGAN_x4 model or the mangaAttempt109 model, look in Deorder's Skyrim Model Google Drive for an already interpolated version.

Fallout 4 weapons was trained using Fallout 4's official hd armor/weapon textures but could be used on other weapon and armor textures.

Normal Map Upscaling

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Normal Maps alsa64 4 GNU GPLv3 Normal Maps 36K 27 128 ? ? Custom (Normal Maps) Normal Maps - Skyrim artifacted
Normal Maps - Skyrim artifacted Deorder 4 Skyrim Normal Maps 145K ? 128 ? ? Skyrim Normal Maps ?

The first one is based on the second one it was trained, with a higher learning rate and insane n_workers and batch_size values. It is meant to replace the old Normal Map model from Deorder, but without adding BC1 compression to your normal maps.

The second one was trained on Skyrim's Normal Maps, including compression artifacts, so it will have to be redone.

Grayscale Upscaling

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Skyrim Alpha Deorder 4 Alpha Channel 105K ? 128 ? ? Alpha Channels from Skyrim ?

Trained to upscale grayscale images, like specular or alpha etc.

Artifact Removal

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
BC1 take 1 alsa64 1 GNU GPLv3 BC1 Compression 100K 2 128 111 1,8K Custom (Photos) Failed Attempts
BC1 take 2 alsa64 1 GNU GPLv3 BC1 Compression 261K 2 128 106 4,7K Custom (Photos / Manga) JPG (0-20%)
BC1 take 3 Noise Aggressive alsa64 1 GNU GPLv3 BC1 Compression 400K 2 128 26 28.985K Custom (just about everything) BC1 take 2
JPG (0-20%) alsa64 1 GNU GPLv3 JPG compressed Images 178K 2 128 52 6.23K Custom (Photos / Manga) JPG (20-40%)
JPG (20-40%) alsa64 1 GNU GPLv3 JPG compressed Images 141K 2 128 42 6.23K Custom (Photos / Manga) JPG (40-60%)
JPG (40-60%) alsa64 1 GNU GPLv3 JPG compressed Images 100K 2 128 31 ~6.5K Custom (Photos / Manga) JPG (60-80%)
JPG (60-80%) alsa64 1 GNU GPLv3 JPG compressed Images 91K 2 128 27 ~6.5K Custom (Photos / Manga) JPG (80-100%)
JPG (80-100%) alsa64 1 GNU GPLv3 JPG compressed Images 162K 2 128 51 ~6.5K Custom (Photos / Manga) BC1 take 1
JPG PlusULTRA twittman 1 JPG compressed Images 130K 1 ? 150 0.937K Custom (Manga) Failed Attempts
Cinepak twittman 1 Cinepak, msvideo1 and Roq 200K 1 128 21 ~8K Custom (Manga) none
DeDither alsa64 1 GNU GPLv3 Dithered Images 127K 2 128 53 4.7K Custom (Photos / Manga) JPG (0-20%)
dither_4x_flickr2k_esrgan, dither_4x_flickr2k_psnr buildist 4 Ordered dithering 280K 16 128 ? 2.64K, ~8K Flickr2K, OST dithered with GIMP none
DeSharpen loinne 1 Oversharpened Images 310K 1 128 48 ~3K Custom (?) Failed Attempts
AntiAliasing twittman 1 Images with pixelated edges 200K 1 128 440 0.656K Custom (?) none

Models to remove compression artifacts.

The BC1 take 2 model is better than my first BC1 model (BC1 take 1), It also might improve edges and tone differences between before and after somewhat. The Dataset was based on the JPG dataset, slightly balanced to contain less manga styled images. Note that BC1 compression is also used for the RGB channel in BC3. BC1=DXT1, BC3=DXT5. Do not use any of them for uncompressed textures.

JPG gets compressed witch a Quality Percentage between 0 and 100. So depending on how bad your JPEGs are compressed, choose the model of your choice. You can use ImageMagick to guess the Quality percentage, but keep in mind that it might be wrong, since the image might have been resaved.

The Cinepak model removes movie compressions artifacts from older video compression methods like Cinepak, msvideo1 and Roq.

Dithering is an older compression method, where the amount of color gets reduced, if your image has few colors or banding try the Dedither model.

Ordered dithering is a less common form of dithering that results in distinctive checkerboard/crosshatch patterns, which are misinterpreted as texture by models not trained on it. It's often used on GIFs because the pattern is stable between frames. For the 4x model, start with the ESRGAN model, and interpolate with the PSNR model if the result is too sharp.

The DeSharpen model was made for rare particular cases when the image was destroyed by applying noise, i.e. game textures or any badly exported photos. If your image does not have any oversharpening, it won't hurt them, leaving as is. In theory, this model knows when to activate and when to skip, also can successfully remove artifacts if only some parts of the image are oversharpened, for example in image consisting of several combined images, 1 of them with sharpen noise. It is made to remove sharpen noise, particulary made with Photoshop "sharpen" or "sharpen more" filters OR Imagemagick's -sharpen directive with several varying parameters of Radius and Sigma, from subtle 0.3x0.5 to something extreme like 8x2, somewhere about that.

AntiAliasing is for smoothing jagged edges in images and textures.

Pretrained models for different scales:

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
1xESRGAN victorca25 1 Pretrained model 1 128 65k Combination of DIV2K, Flickr2K and GOPRO RRDB_ESRGAN_x4.pth
2xESRGAN victorca25 2 Pretrained model 4 128 65k Combination of DIV2K, Flickr2K and GOPRO RRDB_ESRGAN_x4.pth
4xESRGAN victorca25 4 Pretrained model 8 128 65k Combination of DIV2K, Flickr2K and GOPRO RRDB_ESRGAN_x4.pth
8xESRGAN victorca25 8 Pretrained model 16 128 65k Combination of DIV2K, Flickr2K and GOPRO RRDB_ESRGAN_x4.pth
16xESRGAN victorca25 16 Pretrained model 16 128 65k Combination of DIV2K, Flickr2K and GOPRO RRDB_ESRGAN_x4.pth

These models were transformed from the original RRDB_ESRGAN_x4.pth model into the other scales, in order to be used as pretrained models for new models in those scales.

More information can be found here.

Others

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
normal generator LyonHrt 1 Difuse to Normal 215K 1 128 45 4.536K Custom (?) none

The model was trained on pairs of diffuse textures and normal maps.

Other Sources:

Cartoon Painted Models

PPON Models

Name Author Scale License Purpose Iterations Batch Size HR Size Epoch Dataset Size Dataset Pretrained Model
Pixie victorca25 4 Pixel art (and some cartoon) Upscaling 80K 8 192 Drawings downscaled and degraded on the fly PPON
xBRZ+ victorca25 4 Pixel art Upscaling 60K 8 128 xBRZ images (phase 1 and 2) and drawings downscaled and degraded on the fly (phase 3) Pixie PPON

License:

  • GNU GLPv3:
    • You can't sell the model under that license
    • If you modify, interpolate or use the model as a pretrained model for your own model and share results of your resulting model, it will have to be under the same license, meaning that you can't sell it.
    • You have to state that you used the model and its author for your results.
    • You have to state any changes you made to the model.
    • There are other points, but those are the main ones.

In addition to that all models by:

  • alsa/alsa64

have the following additional restriction:

  • You can't sell results generated with a model using that license.