6 AI Upscale Models for Anime Illustrations Compared: UltraSharpV2 vs. IllustrationJaNai_V3

image1787962729 (1600×1600)
  • AI upscale models correct contours
  • ESRGAN makes bold modifications; DAT2 stays faithful to the original
  • Use models in safetensors format

Introduction

Hello, this is Easygoing. In this article, I’d like to compare AI upscale models that can easily enhance image quality.

Anime illustration of a purple-haired girl smiling against a waterfall background
Today’s topic: AI upscale models

I’ve also converted AI upscale models—which are usually released in pickle format—into the safer and potentially higher-performance safetensors format and published them, so I’ll introduce those as well.

AI Upscale Models Correct Details

In a previous article, I explained that there are two ways to increase image resolution: pixel interpolation using mathematical formulas and methods that use AI upscale models.

About AI Upscale Models

AI upscale models can perform high-quality upscaling by correcting elements like contours—which tend to break down with mathematical methods—while increasing resolution.

Comparison Between Mathematical Methods and AI Upscaling

Comparison of Lanczos method and RealESRGAN_Anime_6B: Lanczos Comparison of Lanczos method and RealESRGAN_plus_anime_6B: RealESRGAN_plus_anime_6B
Lanczos | RealESRGAN_plus_anime_6B

Strong Magnification

Strong magnification comparison. Left Lanczos shows blurry eyes, right AI upscale has clear contours

The illustration processed with the AI upscale model on the right has reduced noise in the hair and eyes and clearer contours compared to the Lanczos method on the left.

There are many types of AI upscale models, but this time I’ll compare six models that are particularly suitable for anime illustrations.

The Comparison Baseline Is the Lanczos Method

The comparison will be conducted using the following procedure.

Baseline Image

First, create a baseline image by enlarging the original image to 2× resolution using the Lanczos method.

flowchart LR
A1(Original Image<br>1024 x 1024)
B1(Lanczos Image<br>2048 x 2048)
A1--Lanczos x2-->B1

AI Upscaled Images

Next, generate comparison images using AI upscale models.

flowchart LR
A1(Original Image<br>1024 x 1024)
B1(4096 x 4096)
C1(AI Upscaled Image<br>2048 x 2048)
A1--AI Upscale x4-->B1
B1--Lanczos x0.5-->C1

All AI upscale models used this time have a fixed 4× scale factor, so they are first enlarged 4× and then reduced by 0.5× with the Lanczos method to match the baseline image resolution.

AI Upscale Models Being Compared

The six AI upscale models compared this time are as follows.

gantt
    title AI Upscale Models
    dateFormat YYYY-MM-DD
    axisFormat %Y

    section Nmkd
    YandereNeoXL_ESRGAN               : 2021-01-26, 2026-08-29

    section Tencent ARC Lab
    RealESRGAN_plus_anime_6B          : 2021-08-31, 2026-08-29

    section Kim2091
    UltraSharp_ESRGAN                 : 2021-10-27, 2026-08-29
    UltraSharpV2_DAT2                 : 2025-05-23, 2026-08-29

    section the-database
    IllustrationJaNai_V1_ESRGAN       : 2024-02-23, 2026-08-29
    IllustrationJaNai_V3_denoise_DAT2 : 2025-11-26, 2026-08-29
Model Name Release Month Architecture License
YandereNeoXL_ESRGAN Jan 2021 ESRGAN WTFPL
RealESRGAN_plus_anime_6B Aug 2021 ESRGAN BSD 3-Clause
UltraSharp_ESRGAN Oct 2021 ESRGAN CC-BY-NC-SA-4.0
UltraSharpV2_DAT2 May 2025 DAT2
IllustrationJaNai_V1_ESRGAN Feb 2024 ESRGAN
IllustrationJaNai_V3_denoise_DAT2 Nov 2025 DAT2
  • ESRGAN (2018–): Fast, aggressively modifies the original illustration
  • DAT2 (2023–): Faithful to the original illustration, but heavier processing

The most widespread format among AI upscale models is ESRGAN, which aims to improve quality by aggressively modifying the original illustration. In contrast, the DAT2 format that appeared in 2023 performs upscaling that is faithful to the original, but takes longer to process than ESRGAN models.

Actual Comparisons

Let’s compare actual illustrations. In the comparison images, the left side shows the illustration generated by the AI upscale model, and the right side shows a color difference map compared to the baseline image, with differences emphasized 4× for clarity.

1. High-Quality Anime Illustration

First, let’s upscale a high-quality anime illustration.

Baseline Image

Baseline image 1 of high-quality anime illustration. Purple-haired girl in kimono standing in a Japanese-style room
mellow_pencil-XL_clear_v2 → Krea-2-Turbo_clear

YandereNeoXL_ESRGAN

YandereNeoXL_ESRGAN upscale result for image 1 YandereNeoXL_ESRGAN upscale difference map (4× emphasized) for image 1
AI upscaled image | Color difference map (4× amplified)

RealESRGAN_plus_anime_6B

RealESRGAN_plus_anime_6B upscale result for image 1 RealESRGAN_plus_anime_6B upscale color difference map (4× emphasized) for image 1

UltraSharp_ESRGAN

UltraSharp_ESRGAN upscale result for image 1 UltraSharp_ESRGAN upscale color difference map (4× emphasized) for image 1

UltraSharpV2_DAT2

UltraSharpV2_DAT2 upscale result for image 1 UltraSharpV2_DAT2 upscale color difference map (4× emphasized) for image 1

IllustrationJaNai_V1_ESRGAN

IllustrationJaNai_V1_ESRGAN upscale result for image 1 IllustrationJaNai_V1_ESRGAN upscale color difference map (4× emphasized) for image 1

IllustrationJaNai_V3_denoise_DAT2

IllustrationJaNai_V3_denoise_DAT2 upscale result for image 1 IllustrationJaNai_V3_denoise_DAT2 upscale color difference map (4× emphasized) for image 1

Image1 result

image1 Time (sec) MAE_similarity SSIM_similarity Red Green Blue Brightness
YandereNeoXL_ESRGAN 13.3 99.0 % 99.6 % 0.0 % 0.0 % 0.1 % 0.0 %
RealESRGAN_plus_anime_6B 7.8 98.4 % 99.5 % -0.3 % -0.4 % -0.5 % -0.4 %
UltraSharp_ESRGAN 13.3 98.8 % 99.6 % -0.2 % 0.1 % -0.4 % -0.1 %
UltraSharpV2_DAT2 35.5 99.1 % 99.8 % -0.2 % -0.1 % -0.2 % -0.2 %
IllustrationJaNai_V1_ESRGAN 12.9 99.1 % 99.7 % -0.2 % -0.2 % -0.2 % -0.2 %
IllustrationJaNai_V3_denoise_DAT2 34.7 99.3 % 99.8 % -0.2 % -0.2 % -0.2 % -0.2 %

First, a common characteristic of AI upscale models is that they tend to lower overall brightness and emphasize blacks compared to mathematical methods. Looking at the difference maps and the MAE_similarity and SSIM_similarity values (similarity to the baseline image) in the table, we can see that DAT2-based models make fewer changes from the baseline image than ESRGAN-based models.

Anime illustration of a purple long-haired girl smiling against a waterfall background
ESRGAN models tend to strongly overpaint local areas

Among ESRGAN models, YandereNeoXL_ESRGAN is a relatively conservative model that stays close to the original, while RealESRGAN_plus_anime_6B strongly redraws local areas, especially around character contours.

The IllustrationJaNai series is distinctive in that it adjusts noise and colors across the entire illustration in addition to contours, while the UltraSharp series sits in a well-balanced middle ground among the models compared this time.

2. Raw SDXL Output Illustrations

Next, let’s compare with raw SDXL outputs—the type of illustration where AI upscale models are used most frequently.

Baseline Image

Baseline image 2 of raw SDXL illustration. Purple-haired girl against a waterfall background
mellow_pencil-XL_clear_v2

YandereNeoXL_ESRGAN

YandereNeoXL_ESRGAN upscale result for image 2 YandereNeoXL_ESRGAN upscale color difference map (4× emphasized) for image 2

RealESRGAN_plus_anime_6B

RealESRGAN_plus_anime_6B upscale result for image 2 RealESRGAN_plus_anime_6B upscale color difference map (4× emphasized) for image 2

UltraSharp_ESRGAN

UltraSharp_ESRGAN upscale result for image 2 UltraSharp_ESRGAN upscale color difference map (4× emphasized) for image 2

UltraSharpV2_DAT2

UltraSharpV2_DAT2 upscale result for image 2 UltraSharpV2_DAT2 upscale color difference map (4× emphasized) for image 2

IllustrationJaNai_V1_ESRGAN

IllustrationJaNai_V1_ESRGAN upscale result for image 2 IllustrationJaNai_V1_ESRGAN upscale color difference map (4× emphasized) for image 2

IllustrationJaNai_V3_denoise_DAT2

IllustrationJaNai_V3_denoise_DAT2 upscale result for image 2 IllustrationJaNai_V3_denoise_DAT2 upscale color difference map (4× emphasized) for image 2

Image2 result

image2 Time (sec) MAE_similarity SSIM_similarity Red Green Blue Brightness
YandereNeoXL_ESRGAN 12.6 99.2 % 99.7 % 0.0 % 0.0 % 0.2 % 0.0 %
RealESRGAN_plus_anime_6B 7.9 98.5 % 99.5 % -0.1 % -0.2 % 0.0 % -0.2 %
UltraSharp_ESRGAN 13.1 98.7 % 99.5 % -0.3 % 0.2 % 0.0 % -0.1 %
UltraSharpV2_DAT2 34.2 99.2 % 99.7 % -0.1 % 0.0 % -0.2 % -0.1 %
IllustrationJaNai_V1_ESRGAN 12.9 99.2 % 99.7 % -0.1 % -0.1 % -0.1 % -0.2 %
IllustrationJaNai_V3_denoise_DAT2 34.3 99.4 % 99.8 % -0.2 % -0.1 % -0.1 % -0.2 %

In this illustration, the areas each model corrects are clearly different. RealESRGAN_plus_anime_6B and UltraSharp_ESRGAN models react strongly to edge areas such as the lines of the waterfall.

On the other hand, UltraSharpV2_DAT2 and the IllustrationJaNai series do not react to the waterfall lines and instead focus on overall illustration correction and noise removal.

Because AI upscale models respond differently depending on the environment they were trained in, the optimal model varies from illustration to illustration.

3. Photorealistic Illustrations

Finally, let’s compare with photorealistic illustrations.

Baseline Image

Baseline image 3 of photorealistic illustration. Woman smiling in front of an aquarium
Krea-2-Turbo

YandereNeoXL_ESRGAN

YandereNeoXL_ESRGAN upscale result for image 3 YandereNeoXL_ESRGAN upscale color difference map (4× emphasized) for image 3

RealESRGAN_plus_anime_6B

RealESRGAN_plus_anime_6B upscale result for image 3 RealESRGAN_plus_anime_6B upscale color difference map (4× emphasized) for image 3

UltraSharp_ESRGAN

UltraSharp_ESRGAN upscale result for image 3 UltraSharp_ESRGAN upscale color difference map (4× emphasized) for image 3

UltraSharpV2_DAT2

UltraSharpV2_DAT2 upscale result for image 3 UltraSharpV2_DAT2 upscale color difference map (4× emphasized) for image 3

IllustrationJaNai_V1_ESRGAN

IllustrationJaNai_V1_ESRGAN upscale result for image 3 IllustrationJaNai_V1_ESRGAN upscale color difference map (4× emphasized) for image 3

IllustrationJaNai_V3_denoise_DAT2

IllustrationJaNai_V3_denoise_DAT2 upscale result for image 3 IllustrationJaNai_V3_denoise_DAT2 upscale color difference map (4× emphasized) for image 3

Image3 result

image3 Time (sec) MAE_similarity SSIM_similarity Red Green Blue Brightness
YandereNeoXL_ESRGAN 12.7 99.6 % 100.0 % 0.1 % 0.0 % 0.0 % 0.0 %
RealESRGAN_plus_anime_6B 8.1 98.8 % 99.6 % -0.3 % -0.3 % 0.0 % -0.3 %
UltraSharp_ESRGAN 13.2 99.1 % 99.7 % 0.0 % 0.0 % -0.1 % 0.0 %
UltraSharpV2_DAT2 34.5 99.4 % 99.8 % -0.1 % -0.2 % -0.1 % -0.2 %
IllustrationJaNai_V1_ESRGAN 12.8 99.5 % 99.9 % -0.2 % -0.3 % -0.2 % -0.3 %
IllustrationJaNai_V3_denoise_DAT2 34.0 99.7 % 100.0 % -0.2 % -0.2 % -0.2 % -0.2 %

Since all the models used this time are designed for anime illustrations, YandereNeoXL_ESRGAN and the IllustrationJaNai series were barely able to correct the photorealistic illustration.

RealESRGAN_plus_anime_6B and the UltraSharp series were able to make some corrections to the photorealistic illustration, but the results still differ from the precise corrections they apply to anime illustrations.

When correcting photorealistic illustrations, it’s better to use AI upscale models specialized for photorealistic images.

Overall Evaluation

Now for the overall evaluation.

Upscaling Processing Time

Looking at processing times, the newer DAT2-based models take more than twice as long as the ESRGAN-based models.

Among ESRGAN models, RealESRGAN_plus_anime_6B is especially fast compared to other ESRGAN models. This is because, as the “6B” name indicates, it is a compact model with only 6 layers (typical ESRGAN models have 23 layers).

Characteristics of Each Model

Next, I had Claude analyze the measurement data and all difference maps from this test and create an infographic summarizing the characteristics of each model.

Scatter plot of model characteristics visualizing the balance between fidelity and local/global corrections

According to Claude’s analysis, RealESRGAN_plus_anime_6B performs bold local corrections, while YandereNeoXL_ESRGAN makes corrections that are faithful to the original.

The IllustrationJaNai series applies processing that improves the overall quality of the illustration beyond just contours, while the UltraSharp series delivers well-balanced upscaling overall.

Based on this comparison, IllustrationJaNai series is a good choice when you want to raise the overall quality of an illustration, UltraSharp series when balance is the priority, and RealESRGAN_plus_anime_6B when local corrections are most important.

UltraSharp and IllustrationJaNai Have Non-Commercial Licenses!

When choosing an AI upscale model, checking the license is also important. Among the models compared this time, the UltraSharp series and IllustrationJaNai series are under the CC-BY-NC-SA-4.0 license, which does not permit commercial use.

Model Name Release Month Architecture License
YandereNeoXL_ESRGAN Jan 2021 ESRGAN WTFPL
RealESRGAN_plus_anime_6B Aug 2021 ESRGAN BSD 3-Clause
UltraSharp_ESRGAN Oct 2021 ESRGAN CC-BY-NC-SA-4.0
UltraSharpV2_DAT2 May 2025 DAT2
IllustrationJaNai_V1_ESRGAN Feb 2024 ESRGAN
IllustrationJaNai_V3_denoise_DAT2 Nov 2025 DAT2

About Licenses for Image Generation AI Models

Because the UltraSharp series and IllustrationJaNai series cannot be used commercially, please use different models when generating commercial illustrations.

So What Do I Recommend?

My conclusions are as follows:

  • To raise quality through upscaling → IllustrationJaNai_V3_denoise_DAT2
  • Balance-focused → UltraSharp series
  • Fast generation of commercial illustrations → RealESRGAN_plus_anime_6B

Use Models in Safetensors Format!

ComfyUI’s Load Upscale Model node can load both pickle format and safetensors format models.

ComfyUI Load Upscale Model node. Example showing it can load both .pth and .safetensors formats
Can load both .pth and .safetensors formats
  • Pickle format (.pt, .pth): Executable code
  • Safetensors format (.safetensors): Model weights only, efficient loading

Because pickle format is distributed as executable code, the risk of malicious code such as malware being included cannot be completely ruled out. In contrast, safetensors format extracts only the model weights, making it safer than pickle format.

AI upscale models are small, so there is almost no difference in load time between pickle and safetensors formats. From a security perspective, however, it is better to use safetensors format whenever possible.

Repository of models converted from pickle to safetensors

I’ve published a repository containing many older pickle-format models converted to safetensors format, so feel free to try them out.

Summary: Use AI Upscale Models

  • AI upscale models correct contours
  • ESRGAN makes bold modifications; DAT2 stays faithful to the original
  • Use models in safetensors format

In this article, I compared AI upscale models. AI upscale models are convenient tools that make it easy to upscale images, but the way they apply corrections reflects the philosophy of their creators. Trying various models and experiencing the creators’ thinking is a lot of fun.

Anime illustration of a purple long-haired girl smiling amid water droplets against a waterfall background
Each model has its own personality

The Hugging Face repository mentioned above publishes many models beyond the six introduced here.

The latest IllustrationJaNai_V3 series also includes various variants such as a detail series that enhances details and a lighter FDAT format.

Finding the AI upscale model that best matches your own illustrations through trial and error is very enjoyable, so I encourage everyone to try different ones and discover your favorites.

Thank you for reading to the end!