Releases: Kim2091/Kim2091-Models
1x-AniRestore_TFDAT
@news
1x-AniRestore
Scale: 1
Architecture: TFDAT
Links: https://github.com/Kim2091/Kim2091-Models/releases/tag/1x-AniRestore
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Anti-aliasing, Compression Removal, Debanding, Deblur, Dehalo, Restoration
Subject: Anime, Cartoon, Video Frame
Input Type: Videos
Date: 2/5/26
Size:
I/O Channels:
Dataset: Dataset consisting of Spongebob, Courage, A-KO, Robot Carnival, and Slayers
Dataset Size:
OTF (on the fly augmentations): No
Pretrained Model: Private Spongebob model
Iterations: ~150k total
Batch Size: 16
GT Size: 96
Description: This model has been in development for about 5 months now. Progress started on it quite a while ago with the beginning of TSPAN. leobby and I developed a dataset to restore Spongebob Season 1 and trained models on TSPAN, TSPANv2, and finally TFDAT. TFDAT provided the required level of restoration both temporally and in terms of general detail.
I repurposed part of that Spongebob dataset to make this model, Zarxrax shared his Courage dataset (thank you!), and Bendel shared a really impressive dataset to improve this model. Bendel's contributions allowed this model to dehalo effectively and also fix color bleed! And Zarxrax's dataset improved the model's ability to handle rainbows. Without their contributions, this model wouldn't be anywhere near as good as it is. I'm really happy to have worked with them on this.
This model can handle: Rainbows, Color Bleed, Dot Crawl, general shimmering, blur, and some amounts of noise. Note that if the show you intend to restore has thicker lines, this model likely won't work well on it.
To use this model, download and install my program Vapourkit. Then just install the plugins in the top left, import the model, and you're good to go! Make sure you're in TensorRT mode!
Showcase: https://slow.pics/c/tZ7xglRT
AniRestore.mp4
1x-DeBink_TFDAT
@news
1x-DeBink_TFDAT
Scale: 1
Architecture: TFDAT
Author: kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, Debanding, Denoise, Restoration, Texture Generation
Subject: Anime, Cartoon, CGI, Video Frame
Input Type: Videos
Date: 1/24/26
Size:
I/O Channels:
Dataset: Custom DeBink Dataset
Dataset Size: 12150 pairs, 2430 scenes
OTF (on the fly augmentations): No
Pretrained Model:
Iterations: ~200k total
Batch Size: 8
GT Size: 64
Description: The first model I ever trained and released here was DeBink. It was an ESRGAN model that attempted to remove Bink compression. It worked reasonably well for what it was, but it had major issues.
This fixes all of them! Trained on a VSR architecture, it's able to reconstruct detail from the surrounding frames and restore details and information that were missing from the compression blocks. It looks very natural and smooth.
The model also fixes the issues that Bink videos usually have, which is their washed out appearance. Most Bink videos I tested with had this issue.
To use this model, you'll need to download the FP32 model, install Vapourkit, disable DirectML mode in the top left, select "Import Model", and then select TSPAN/Dynamic/FP32. Build it and you're good to go! It can be used in DirectML mode too, but it is very slow.
Showcase:
b.mp4
4x-UltraSharpV2
<@&560103931204861954>
UltraSharpV2
Scale: 4
Architecture: DAT2
Links: https://ko-fi.com/s/4b3245cfe5
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, Dehalo, Denoise, General Upscaler, JPEG, Restoration
Subject: AI Generated, Anime, Cartoon, CGI, Faces, Screenshots, Textures, Photography, Text
Input Type: Images
Date: 5/23/25
Dataset: Private dataset
Dataset Size:
OTF (on the fly augmentations):
Pretrained Model:
Iterations:
Batch Size:
GT Size:
Description: This model has been in the works for a very very very long time. I spent years creating my own private dataset just to train this model. It can handle realistic images, anime, cartoons, and quite a bit more without much hassle. It also works amazingly on illustrations and artwork. This is easily the best model I've ever trained.
I hope you all like it.
The Lite model is based on the RealPLKSR architecture, rather than DAT2.
And please consider donating! I'm going through a rough time right now & any support helps a LOT
Showcase: https://slow.pics/s/RORQ4R0U?image-fit=contain
Go to my Ko-fi page to get the model for free!
4x-PBRify_UpscalerV4
@news
4x-PBRify_UpscalerV4
Scale: 4
Architecture: DAT2
Links: https://github.com/Kim2091/Kim2091-Models/releases/tag/4x-PBRify_UpscalerV4
Author: Kim2091
License: CC0
Purpose: Compression Removal, General Upscaler, Restoration
Subject: Game Textures
Input Type: Images
Date: 5/19/25
Size:
I/O Channels:
Dataset: Custom texture dataset
Dataset Size:
OTF (on the fly augmentations): No
Pretrained Model: 4x-PBRify_UpscalerDAT2_V1
Iterations: 200k
Batch Size: 8
GT Size: (LQ size) 48
Description: A new version of the main PBRify upscaling model. The PBRify Upscaler series of models are meant to take existing game textures from older 2000s era games, and upscale them to usable quality.
V4 significantly improves detail over the previous V3 model. It is slower, as it's based on the DAT2 architecture, however the results are very much worthwhile imo.
Showcase: https://slow.pics/c/vMGFFfFh
2x-GameUp_TSCUNet
@news
2x-GameUp_TSCUNet
Scale: 2
Architecture: TSCUNet
Links:
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, General Upscaler, Restoration
Subject: Video Frame
Input Type: Videos
Date: 3/26/25
Size:
I/O Channels:
Dataset: Custom game dataset
Dataset Size: 11150 Pairs, 223 Scenes
OTF (on the fly augmentations): No
Pretrained Model: 2x_eula_anifilm_2022_vsr
Iterations: 160k
Batch Size: 8
GT Size: (LQ size) 64
Description: This is my first video model! It's aimed at restoring compressed video game footage, like what you'd get from Twitch or Youtube. I've attached an example below.
It's trained on TSCUNet using lossless game recordings, and degraded with my video destroyer. The degradations include resizing, and H264, H265, and AV1 compression.
IMPORTANT: You cannot use this model with chaiNNer or any other tool. You need to use this.
You just run test_vsr.py after installing the requirements. Use the example command from the readme.
You can also use the ONNX version of the model with test_onnx.py
If you want to train a TSCUNet model yourself, use traiNNer-redux. I've included scripts in the SCUNet repository to convert your own models to ONNX if desired.
Showcase: Watch in a Chrome based browser: https://video.yellowmouse.workers.dev/?key=Fvxw482Nsv8=
2x-GameUpV2_TSCUNet Set
<@&560103931204861954>
2x-GameUpV2_TSCUNet + Small
Scale: 2
Architecture: TSCUNet + TSCUNet Small
Links: https://github.com/Kim2091/Kim2091-Models/releases/tag/2x-GameUpV2_TSCUNet
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, General Upscaler, Restoration
Subject: Video Frame
Input Type: Videos
Date: 3/28/25
Size:
I/O Channels:
Dataset: Custom game dataset
Dataset Size: 16310 Pairs, 395 Scenes
OTF (on the fly augmentations): No
Pretrained Model: 2x-GameUpV1_TSCUNet
Iterations: 180k
Batch Size: 8
GT Size: (LQ size) 64
2x-GameUpV2_TSCUNet Small Description: Just a quick release on what I've been calling TSCUNet Small. It's just the original arch but with nb cut from 2 to 1. It improves performance by ~40-60%, and it wasn't too hard to get the quality to match the full V2 model. This should make it more accessible and run on more PCs :)
No comparisons this time though, sorry. It's pretty much identical visually to V2.
2x-GameUpV2_TSCUNet Description: The second version of GameUp. It reduces noise a LOT in the final output, and looks a lot better as a whole. I expanded the dataset with more game recordings, with more varied scenery, then prepared with my video destroyer. The degradations include resizing, and H264, H265, and AV1 compression.
IMPORTANT: You cannot use this model with chaiNNer or any other tool. You need to use this tool. There's a GUI and also TRT support separately.
To train one yourself, use traiNNer-redux.
NOTE: ONNX models were updated on April 7 2025 in conjunction with an update to the SCUNet Repository. The old ONNX models are no longer compatible with the inference code, and the new ones support TensorRT! Look here for more info: https://github.com/Kim2091/SCUNet/blob/main/tensorrt/README.md
Showcase: Watch in a Chrome based browser: https://video.yellowmouse.workers.dev/?key=z6MbPmhuML8=
V1 vs V2: https://video.yellowmouse.workers.dev/?key=aE4oxj3lydo=
2x-AnimeUp
2x-AnimeUp
Scale: 2
Architecture: TSPAN
Links: https://github.com/Kim2091/Kim2091-Models/releases/tag/2x-AnimeUp
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, Debanding, General Upscaler, JPEG, Restoration
Subject: Anime, Cartoon, Video Frame
Input Type: Videos
Date: 10/3/25
Size: 48 Features
I/O Channels: 3(RGB)->3(RGB)
Dataset: Custom dataset made with video destroyer
Dataset Size: 3000
OTF (on the fly augmentations): No
Pretrained Model: 1x-SuperScale
Iterations: 170k
Batch Size: 8
GT Size: 128
Description: This is the first model for my new architecture TSPAN! TSPAN is a VSR architecture based on SPAN. It is temporally stable, yet still very fast. On a 4090, 480p input can be upscaled by 2x at ~140 FPS with TensorRT.
As for the model itself, it's been trained on MPEG-2 and H264 compression and various resizing filters. It's a pretty simple model, but it performs very well on its intended input (modern anime that's been moderately compressed).
Important: To use this model, you must use my code published here. It is not compatible with tools like chaiNNer as of right now.
To train a TSPAN model, you should use this code: https://github.com/Kim2091/traiNNer-redux-1. Just follow the readme, then use the TSPAN config. You must also prepare a video dataset, I'd recommend learning using video destroyer.
Please note that the model does not actually shift colors! I don't know why the comparisons have color shifting :(
Video Demo: (Watch in Chrome based browsers)
https://kaizoku.pw/c/nRCyeU2EvU9D92CubqAC
Frame Comparison:
https://slow.pics/s/8Vxe2KDc?image-fit=contain
Animated Comparison:
1x-SwatKats_DIS_Balanced
1x-SwatKats_DIS_Balanced
Scale: 1
Architecture: DIS
Links:
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, Debanding, Denoise, JPEG, Pretrained, Restoration
Subject: Anime, Cartoon, Video Frame
Input Type: Images
Date: 11/26/25
Size: Max (32 features, 12 blocks)
I/O Channels: 3(RGB)->3(RGB)
Dataset: Modified SwatKats dataset
Dataset Size:
OTF (on the fly augmentations): No
Pretrained Model: None
Iterations: ~150k total
Batch Size: 48
GT Size: 128
Description: This is the first model trained on my new ultra lightweight architecture, DIS. From my tests, the DIS_Balanced variant of it is nearly as capable as a full size Compact model while being about 25% faster in TensorRT! This model can be used in pretty much any community tool like chaiNNer, Vapourkit, VideoJaNai, etc.. Maybe even fast enough to be used in real-time!
Benchmarks:
720p, TensorRT FP16 Dynamic engine
1x-SwatKats_DIS_Balanced = 100 FPS
1x-SwatKats_Compact = 75 FPS
This model should hold up compared to the original Compact model. I tested it on many images, and there weren't many places where it differed. In most cases it was just messing up in different areas than Compact did, but both messed up still 😆.
Additionally, I have included a glsl shader for mpv! It's fast enough to be used in realtime on a good amount of hardware. It introduces some slight color shift, but that's the only issue vs the original model.
Showcase: https://slow.pics/c/PThScS2W
SwatKats_DIS_Max.mp4
1x-SuperScale
1x-SuperScale
Scale: 1
Architecture: SPAN, RealPLKSR Small/Tiny
Links: https://github.com/Kim2091/Kim2091-Models/releases/edit/1x-SuperScale
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Anti-aliasing, Restoration
Subject:
Input Type: Images
Date: 5-5-25
Size: Tiny/Small
I/O Channels: 3(RGB)->3(RGB)
Dataset: 8k Dataset V3, Custom Ansel dataset
Dataset Size: A lot
OTF (on the fly augmentations): No
Pretrained Model: 2x_DF2K_Redux_RealPLKSRLayerNorm_450k, 2x_BHI_small_Redux_SPAN_S_1m30k
Iterations: 300k total for SPAN, 100k each for RPLKSR
Batch Size: look at attached configs
GT Size: ^
Description: I was bored, so I did this. This model uses DPID as the scaling algorithm for the HRs. The original images were 8k or 12k. It's significantly sharper than Box/Area scaling, yet does a great job with aliasing. This allows for a very sharp model with minimal artifacts, even on the SPAN version.
The main model is trained on 12k images captured with Nvidia Ansel. It took about 2 days capturing manual 4k and 12k pairs for this model. The 4k captures were used as the LR, the 12k captures were resized to 4k with DPID with randomized lambda values, then trained on as HRs.
The Alt model is trained exclusively on 8k images from my 8k dataset, resized to 4k with dpid. This provides a clearer result with less noise, but it doesn't handle long edges well at all.
Thanks to CF2lter for advice on preparing the dataset, and umzi2 for creating the rust version of DPID.
Showcase: https://slow.pics/c/TCyqje9K
1x-FullTone
FullTone
Scale: 1
Architecture: GateRV3
Links: https://github.com/Kim2091/Kim2091-Models/releases/tag/1x-FullTone
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Compression Removal, Restoration
Subject: Photography, Text
Input Type: Images
Date: 6/12/25
Dataset: Dataset created by @saturnized, refined by myself
Dataset Size: 1300 tiles
OTF (on the fly augmentations): No
Pretrained Model: 1x_umzi_digital_decompress_gaterv3_1
Iterations: ~400k total
Batch Size: 8
GT Size: 64
Description: This is a model that started development back in ~2022. @saturnized shared halftone scans & HQ pairs with me in hopes a model could be trained on it. Many years later we've finally gotten some good results! The HRs were almost entirely reworked.
This model works best on 1200 dpi color halftone scans (uncorrected), and 600dpi BW halftone scans. If your halftone patterns aren't being removed, try resizing the image to ~50% scale. Not intended for use on comics, illustrations, or similar.
The Alt model was trained on resized scans to allow below 1200 DPI to work better, however full res scans suffer.
Unfortunately this model isn't compatible with chaiNNer yet, GateRV3 hasn't been added to spandrel. I will update this when that happens. For now, the only way to run this model is through traiNNer-redux.
Showcase:
