Collection of models for use in restoration and upscaling of the TV series "Where on Earth is Carmen Sandiego?" from the 1990s. While primarily consisting of cel-based animation, the show also incorporates live-action (including stock footage and scenes involving "The Player"), 3D CGI (primarily used for the C5 Corridor), and 2D computer animation. The models are divided into two groups:

* Deinterlace Fix
* Upscaling

The deinterlacing fox models are primarily designed for the 2012 Mill Creek release, where a number of episodes have some rather harsh forced deinterlacing, which the models attempt to repair. These models may also be useful for improving picture quality in some scenes (primarily in seasons 3 and 4) that used deinterlacing as part of compositing multiple layers of images, and for improving the results of deinterlacing by QTGMC or other processes.

Speaking of deinterlacing, for episodes in 480i, I recommend using QTGMC to deinterlace them to 60 (59.97) fps, the frame rate of the live-action and (some) C5 Corridor footage. For the pre-deinterlaced episodes, you're going to be stuck at 30 fps since the other-frame information has already been removed. Despite this series being primarily cel-based animation, I **don't** recommend using 24 fps as it tends to make higher-framerate portions look odd.

The upscaling models work pretty much how you would expect.

Each model has a small and large variant. The large variants are based on the OmniSR architecture, and tend to produce better results (marginally so in the case of the deinterlacing fix), but are several times slower than the small models, which use the compact architecture.

Where On Earth Is Carmen Sandiego?

The "Complete Series" release of "Where on Earth is Carmen Sandiego?" by Mill Creek has a number of episodes that suffer from some rather harsh deinterlacing. This model attempts to restore the footage to something closer to a more advanced deinterlacing job. It may also be useful on episodes (primarily found in seasons 3 & 4) that were produced with advanced compositing where certain layers used similar forced deinterlacing techniques, and for cleaning up artifacts left by Yadif, QTGMC, etc.

Revision History:
* v1.1.0: Added data to improve handing of certain situations, particularly in season 3.
* v1.0.0: Initial release.

Where On Earth-Deinterlace Fix-Large

Where On Earth-Deinterlace Fix-Small

Upscale 90s and early 2000 American cartoons

90s Cartoon v1

Purpose: Anime/Pretrained

A 2x Ultra Lite model coming in under 8MB. Trained with over 15 sets of LRs ranging in a wide amount of issues. Handles Rainbows, Dot Crawl, MPEG/H.264 Compression, and may even assist in removing halos, and fixing blurriness in certain cases. This is my first public model for everyone. Best when used on old anime that is grainy. I can't say what anime it's best suited for as I have tried multiple series, and have found it does a good job on most all the tests. I wouldn't say use this for Western Animation, but it may work. I have done a few tests that I have shown in the upscale results, but that was chained with other models to achieve such a result. This model is meant to retain the more natural look of a series. There is a color shift on the end result, not drastic, but still noticable. I figure you should fix any color issues in post that way to give a more polished upscale. Big thanks to @SaurusX for the model name, and just helping out in general with anything.

AnimeClassics UltraLite

Purpose: Anime

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>


![1729798851 305732](https://github.com/user-attachments/assets/4d7aaf33-5a39-4f75-8a05-90b90a693e49)

AnimeSharpV3

AnimeSharpV3_RCAN

[Link to Github Release with more infos to the process](https://github.com/Phhofm/models/releases/tag/2xAoMR_mosr)

2xAoMR_mosr  
Scale: 4  
Architecture: [MoSR](https://github.com/umzi2/MoSR)  
Architecture Option: [mosr](https://github.com/umzi2/MoSR/blob/95c5bf73cca014493fe952c2fbc0bdbe593da08f/neosr/archs/mosr_arch.py#L117)  

Author: Philip Hofmann  
License: CC-BY-0.4  
Purpose: A 2x mosr upscaling model for game textures  
Subject: Game Textures  
Input Type: Images  
Release Date: 21.09.2024 (dd/mm/yy)  

Dataset: Game Textures from Age of Mythology: Retold  
Dataset Size: 13'847  
OTF (on the fly augmentations): No  
Pretrained Model: [4xNomos2_hq_mosr](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_mosr)  
Iterations: 510'000  
Batch Size: 4  
Patch Size: 64  

## Description:   
In short: A 2x game texture mosr upscaling model, trained on and for (but not limited to) Age of Mythology: Retold textures.

Since I have been playing Age of Mythology: Retold (casual player), I thought it would be interesting to train an single image super resolution model on (and for) game textures of AoMR, but this model should be usable for other game textures aswell.
This is a 2x model, since the biggest texture images are already 4096x4096, I thought going 4x on those would be overkill (also there are already 4x game texture upscaling models, so this model can be used for similiar cases where 4x is not needed).

Model Showcase:
[Slowpics](https://slow.pics/s/IlkmsToH)

2xAoMR_mosr

This model is designed to upscale the standard definition digital era of the Pokémon anime, which runs from late season 5 (Master Quest) to early season 12 (Galactic Battles). During this time, the show was animated digitally in a 4:3 ratio. This process was also used for Mewtwo Returns, most of Pokémon Chronicles, and the Mystery Dungeon specials.

Advice/Known Limitations:
* This OmniSR model can occasionally produce black frames when run in fp16 mode. This seems to be more common in the TPCi era (seasons 9 and later). The issue is sporadic enough that it probably makes sense to do a first pass in fp16, then re-upscale any affected shots in fp32.
* I recommend using QTGMC on a preset of "Slow" or slower for deinterlacing. While the show is primarily animated at 12/24 fps, some elements like backgrounds are animated at a full 60i.
* The model is not great at handling fonts, particularly the italicized text in the episode credits. This is despite including font images in the training data,.

Digital Pokémon-Large

This model is designed to upscale the standard definition digital era of the Pokémon anime, which runs from late season 5 (Master Quest) to early season 12 (Galactic Battles). During this time, the show was animated digitally in a 4:3 ratio. This process was also used for Mewtwo Returns, most of Pokémon Chronicles, and the Mystery Dungeon specials.

Advice/Known Limitations:
* I recommend using QTGMC on a preset of "Slow" or slower for deinterlacing. While the show is primarily animated at 12/24 fps, some elements like backgrounds are animated at a full 60i.
* The model is not great at handling fonts, particularly the italicized text in the episode credits. This is despite including font images in the training data,.

Digital Pokémon-Small

A 2x Compact anime upscale model that handles AVC (h264) degradation. Applied h264 crf 20-28 degradation together with bicubic, bilinear, box and lanczos downsampling on musl's HFA2k dataset with Kim's dataset destroyer.

2xHFA2kAVCCompact

A 2x EDSR M anime upscale model that handles AVC (h264) degradation. Applied h264 crf 20-28 degradation and bicubic, bilinear, box and lanczos downsampling on musl's HFA2k dataset with Kim's dataset destroyer.

2xHFA2kAVCEDSR_M

The second released community trained model on the OmniSR network. Trained with multiscale discriminator to fix decreased output brightness occuring with OmniSR. 2x anime upscale that handles AVC (h264) compression since h264 crf 20-28 degradation together with bicubic, bilinear, box and lanczos downsampling has been applied on musl's HFA2k dataset with Kim's dataset destroyer.

2xHFA2kAVCOmniSR

2x SRFormer_light anime upscale model that handles AVC (h264) compression since h264 crf 20-28 degradation together with bicubic, bilinear, box and lanczos downsampling has been applied on musl's HFA2k dataset with Kim's dataset destroyer for training.


If you want to run this model with chaiNNer (or another application) you need to use the onnx files with an onnx upscale node. All onnx conversions can be found in the [onnx folder on my repo](https://github.com/Phhofm/models/tree/main/2xHFA2kAVCSRFormer_light/onnx).

Example 1: https://imgsli.com/MTkxMTQz

Example 2: https://imgsli.com/MTkxMTQ0

2xHFA2kAVCSRFormer_light

SPAN model for live action film and digital video. The main goal is to fix/reduce common video quality problems while maintaining fidelity. I tried the existing video-focused models and they all denoise or cause colour shifts so I decided to train my own.

The model is trained with compression (JPEG, MPEG-4 ASP, H264, VP9, H265), chroma subsampling, blurriness from multiple scaling, uneven horizontal and vertical resolution, oversharpening halos, bad deinterlacing jaggies, and onscreen text. It is not trained to remove noise at all so it preserves details in the source well. To prevent colour/brightness shifts, I used consistency loss in neosr. I had to modify consistency loss to use a stronger blur so it doesn't interfere with the halo removal.

Limitations:
1. The model has limited ability to see details through heavy grain, but light to moderate grain is fine.
2. The model still does not handle bad deinterlacing perfectly, especially if the source is vertically resized. Fixing bad deinterlacing is not the main goal so it is what it is. Sources that are line-doubled throughout should be descaled back to half height first for best results.
3. The model sometimes oversharpens a little. This is probably because the training data has some oversharpened images.
4. This model generally cannot handle VHS degradation.

More comparisons: https://slow.pics/c/DtDN7gaq

The training config and image degradation scripts used to create training data can be found in https://github.com/jcj83429/upscaling/tree/9332e7d5b07747ff347e5abdc43f8144364de9f7/2xLiveActionV1_SPAN

2xLiveActionV1_SPAN

Upscale modern animation images/videos. Compression, blur, and un-sharpening degradations were used on the dataset. First model I trained and I think it came out pretty well.

ModernSpanimationV1

[Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR)](https://github.com/Francis0625/Omni-SR)

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

OmniSR 2x DF2K

OmniSR 2x DIV2K

Purpose: Compression Removal, Noise Reduction, Line Correction, MPEG2 / LD Artifact Removal

This is my first model release. The model was inspired from a personal project which I have been pursuing for some time now, which this model aims to solve. This model will upscale low resolution hand drawn animation from 1970s to 2000. The colors are retained with effective noise control. Details and Textures are maintained to a good degree considering animations. Color Spills are also corrected depending on the colors. Shades of white and yellows have been difficult. It also makes the lines slightly sharper and thinner. This could be a plus depending on your source. The model is also temporally stable across my tests with little observable issues. 

**Showcase:** 
Images - https://imgsli.com/MjYwNzY1/12/13
Video Sample -  https://t.ly/Jp7-w vs Upscale - https://t.ly/PdsKs

2x Pooh V4

[Swift-SRGAN - Rethinking Super-Resolution for real-time inference](https://github.com/Koushik0901/Swift-SRGAN)

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 2x

Category: VHS Tapes
Purpose: VHS

VHS captures of Film material, but may work on VHS recorded native SD-TV material as well. Also useable for cleaned up source material

VHS upscale and denoise Film

Purpose: VHS Restoration

An advanced VHS recording model designed to enhance video quality by reducing artifacts such as haloing, ghosting, and noise patterns. Optimized primarily for PAL resolution (NTSC might work good as well).

VHS2HD

Upscaling model for 1990s cartoon series "Where on Earth is Carmen Sandiego?" May also be useful for other shows of that era.

Note: Before using this model, you will need to either:
* Deinterlace the episodes (I recommend QTGMC on Slower) to 60 fps (I recommend 60 fps because that's the rate of the live-action footage and some other components).
* Repair the existing deinterlacing via one of the related models for this show.
Which method will depend on the episode and release. See the related deinterlacing models for details.

This show is...complicated. Unlike, say, the Super Mario Bros. Super Show, where live action and animated segments are generally separated into large chunks, Where on Earth is Carmen Sandiego repeatedly switches between live-action, traditional animation, 3D CGI, 2D computer graphics, stock photography, and other sources, sometimes combining them in the same frame. As such, rather than create multiple models for each type, I've created a single model for all footage types.

Known Weaknesses:
* The model performs less well on the live-action (Player) and 3D CGI (C5 Corridor) segments due to a scarcity of usable data in those scenes.
* I don't own the 2025 DVD release, so I can't comment on how well this model works on that footage.

Where On Earth-Upscale-Large

Where On Earth-Upscale-Small

This model is based on a private model by @eula 5600x 3070 named 4x_eula_anifilm_v1_225k. He sent me a copy of the model, and I decided to train a compact model based on it with his permission. This model seems to fix the majority of the issues the original model had while being far faster, it's just a tiny bit softer in some images.

The dataset consists of Dragon Ball movies converted to YUV24 with @sgdisk --zap-all /dev/sda's help to reduce artifacts, then upscaled with ArtClarity and eula_anifilm. LRs are the original frames right from DVD. As a result, this model corrects some color space issues. The 2x model's HRs were downscaled by 50% with Lanczos.

The 2x and 4x models are pretty close in output despite being trained separately. The 2x model is a bit softer overall.

The models in the Real-ESRGAN Compatible folder are the original output from Real-ESRGANs training code for compatibility reasons.

Anifilm Compact (2x)

A 2x model designed for hand-drawn animation of the 80's, 90's and early 2000s.

The Super Mario Bros. Super Show is a hybrid animated and live-action show that debuted in 1989. Each episode featured a cartoon segment sandwiched between live-action segments. The combination of the two formats makes the show very difficult to upscale, and creates the need for multiple model files.

These models were trained using footage from the show itself (except for some title images that were recreated in Photoshop) by finding cases where the same image was shown at different resolutions. For the animated footage, I also did some retouching to reduce haloing, color bleed, and other issues.

A few notes:

- The opening sequence (in particular, the first shot with the giant talking Mario face) is very difficult to upscale due to the picture being degraded from multiple layers of compositing. Ideally, a professional remaster of the series would have access to the individual elements, upscale those, then recomposite them.
- Although live-action models are provided, you will likely get better results from commercial upscaling software (I recommend  the Proteus model in Topaz Video AI).
- Footage must be deinterlaced before processing (I recommend QTGMC for both live-action and animated segments).

[Full Read Me](https://www.mediafire.com/file/brp4c0pvv5s42rw/ReadMe.txt/file)

[Sample Animated Result](https://www.mediafire.com/view/wtieecra7iov35f/Animation_Sample.png/file)

[Sample Live Action Result](https://www.mediafire.com/view/xa7jenxl18xd1y7/Live_Action_Sample.png/file)

Where On Earth-Upscale-Small

Collections that include this model

Similar Models

Architecture	Compact
Scale	2x
Color Mode	RGB
License	MIT Private use Commercial use Distribution Modifications Credit required No Liability & Warranty Disclaimer
Date	2025-08-03
Dataset	Footage from Shout Factory and Mill Creek DVD releases, and some manually-created images.
Dataset size	80
Training iterations	203247
Training batch size	30
Training HR size	64
Training OTF	No