Releases: Kim2091/Kim2091-Models
4x-DAT2_mssim_Pretrain
4x-DAT2_mssim_Pretrain
Scale: 4
Architecture: DAT2
Author: Kim2091
License: CC0
Purpose: Pretrained
Subject:
Input Type: Images
Date: 6-5-24
Size:
I/O Channels: 3(RGB)->3(RGB)
Dataset: UltraSharpV2
Dataset Size:
OTF (on the fly augmentations): No
Pretrained Model: None
Iterations: 50k
Batch Size: 6
GT Size: 128
Description: Simple pretrain for DAT2, trained on CC0 content. "Ethical" model
2x-AnimeSharpV4
@News
2x-AnimeSharpV4
Scale: 2
Architecture: RCAN
Links: Github Release
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Anime
Subject:
Input Type: Images
Date: 1-7-25
Size:
I/O Channels: 3(RGB)->3(RGB)
Dataset: ModernAnimation1080_v3
Dataset Size: 6k
OTF (on the fly augmentations): No
Pretrained Model: 2x-AnimeSharpV3_RCAN
Iterations: 100k RCAN
Batch Size: 8
GT Size: 64
Description: This is a successor to AnimeSharpV3 based on RCAN instead of ESRGAN. It outperforms both versions of AnimeSharpV3 in every capacity. It's sharper, retains even more detail, and has very few artifacts. It is extremely faithful to the input image, even with heavily compressed inputs.
Currently it is NOT compatible with chaiNNer, but will be available on the nightly build soon (hopefully).
Comparisons: https://slow.pics/c/63Qu8HTN
2x-AnimeSharpV3
@News
2x-AnimeSharpV3
Scale: 2
Architecture: ESRGAN & RCAN
Links: Github Release
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Anime
Subject:
Input Type: Images
Date: 10-24-24
Size:
I/O Channels: 3(RGB)->3(RGB)
Dataset: ModernAnimation1080_v3
Dataset Size: 2-3k
OTF (on the fly augmentations): No
Pretrained Model: 4xESRGAN (2x_DF2K_Redux_RCAN_500k.safetensors RCAN)
Iterations: 140k (50k RCAN)
Batch Size: 8
GT Size: 64-128
Description: This release contains an ESRGAN and an RCAN version. Both provide superior quality compared to AnimeSharpV2 in nearly every scenario. It has most of the advantages of the old V2 Sharp models, while not having issues with depth of field.
The RCAN model outperforms the ESRGAN model by a significant margin, with much more consistent generation and overall better detail retention. Currently it is NOT compatible with chaiNNer, but will be available on the nightly build soon (hopefully).
RCAN vs ESRGAN: https://slow.pics/c/Zqgl62Ni
Comparisons: https://slow.pics/c/A2BRSa0U
2x-AnimeSharpV2_Set
@news
2x-AnimeSharpV2 Set
Scale: 2
Architecture: RealPLKSR & MoSR GPS
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Anime
Subject:
Input Type: Images
Date: 10-3-24
Size:
I/O Channels: 3(RGB)->3(RGB)
Dataset: HFA2k modified
Dataset Size: 2-3k
OTF (on the fly augmentations): No
Pretrained Model: 4x_realplksr_mssim_pretrain & 4x-MoSR_GPS_Pretrain
Iterations: 100k & 75k
Batch Size: 6-10
GT Size: 64-256
Description: This is my first anime model in years. Hopefully you guys can find a good use-case for it. Included are 4 models:
- RealPLKSR (Higher quality, slower)
- MoSR (Lower quality, faster)
There are Sharp and Soft versions of both
When to use each:
- Sharp: For heavily degraded sources. Sharp models have issues depth of field but are best at removing artifacts
- Soft: For cleaner sources. Soft models preserve depth of field but may not remove other artifacts as well
Notes:
- MoSR doesn't work in chaiNNer currently
- To use MoSR:
- Use the ONNX version in tools like VideoJaNai
- Update spandrel in the latest version of ComfyUI
The ONNX version may produce slightly different results than the .pth version. If you have issues, try the .pth model.
Comparisons: https://slow.pics/c/4UI20Qlu
1x-UnResizeOnly_RCAN
1x-UnResizeOnly_RCAN
Scale: 1
Architecture: RCAN
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Artifact Removal
Subject:
Input Type: Images
Date: 1-6-25
Size:
I/O Channels: 3(RGB)->3(RGB)
Dataset: UltraSharpV2_Ethical, DigitalArtV3, ModernAnimation1080_v3, Kim2091's 8k Dataset V2
Dataset Size: 13k 512x512 tiles
OTF (on the fly augmentations): No
Pretrained Model: 2x_DF2K_Redux_RCAN_500k.safetensors
Iterations: 95k
Batch Size: 8
GT Size: 64
Description: A version of UnResize trained on RCAN, which is faster and provides better quality than ESRGAN
This model does not remove compression or perform deblurring, unlike the original UnResize models. It only removes scaling artifacts.
I've attached the script I used to create the dataset (it utilizes imagemagick) and the config for traiNNer-redux
1x-BroadcastToStudio_Compact
1x-BroadcastToStudio_Compact
Scale: 1
Architecture: Compact
Author: Kim2091
License: CC BY-NC-SA 4.0
Purpose: Cartoons
Subject: Restoration
Input Type: Images
Date: 12-1-24
Size:
I/O Channels: 3(RGB)->3(RGB)
Dataset: BroadcastToStudio
Dataset Size: 6k
OTF (on the fly augmentations): No
Pretrained Model: 1x-SwatKats_Compact
Iterations: 8k+25k
Batch Size: 8
GT Size: 96
Description: This is a simple retrain of SaurusX's 1x_BroadcastToStudioLite_485k model from a couple years ago. This one is trained on compact, actually has less artifacts, and is significantly faster.
Comparisons: https://slow.pics/c/oGwHyYym
2x-GameScaler_hdsrnet
A model trained on hdsrnet, a new and efficient Image SR architecture. Currently not supported by any software but neosr
1x-GameSmooth_SuperUltraCompact Release
Refer to this folder for additional info: https://github.com/Kim2091/Kim2091-Models/tree/main/1x-GameSmooth_SuperUltraCompact
This model is released under the CC0 license.