3.0.0

V3 release of IllustrationJaNai models for color illustration feature better detail and texture generation, improved correction of line aliasing, better color and contrast accuracy, smoothing for limited 256 color images. V3 includes two variants, each trained on 4 different networks for varying levels of quality and speed, for a total of 8 models. Sample images for all 8 models: https://slow.pics/c/k9Qknnpg?image-fit=cover

• V3detail: all previous versions of IllustrationJaNai were trained to clean halftone noise, which is common for manga covers, but sometimes this also reduces detail and texture in images that don't have any halftone noise. V3detail does not clean any halftone noise, and instead retains as much detail and texture as possible. On some images V3detail may go too far in creating noise and texture, and V3denoise could provide a better result.
• V3denoise: a direct successor to IllustrationJaNai_V2, which cleans halftone noise, but includes all the other improvements of V3 such as better correction of line aliasing, better color and contrast accuracy, smoothing for limited 256 color images. On some images V3denoise may remove too much detail, and V3detail could provide a better result.

For the optimal result for a given image it's best to try both variants and choose the best upscale.

Each variant was trained on 4 different networks; the detail variant was trained on a 5th network HAT L for even higher quality. The denoise variant did not benefit from HAT L as the model became much more noisy than the DAT2 version.

Network	Quality	Time to upscale 764x1200 image in ChaiNNer on RTX 4090
4x HAT L	Best (Detail)	18 seconds (but more than double the VRAM usage vs DAT2, not suitable for lower VRAM GPUs or high resolution images)
4x DAT2	Best (Denoise), Nearly Best (Detail)	15 seconds
4x FDAT XL	Better	4.8 seconds
4x FDAT M	Good	1.8 seconds
2x FDAT M unshuffle	Good (2x)	0.3 seconds
2x SPAN	Good on lightly degraded images (2x)	0.08 seconds

The FDAT architecture is not currently supported in chaiNNer, so any models with FDAT in the name do not work in chaiNNer. Use one of the following options:

• Upscale using the latest release of MangaJaNaiConverterGui, which supports FDAT.
• Patch chaiNNer nightly to support FDAT, download spandrel-fdat.zip and extract it to %APPDATA%\chaiNNer\python\python\Lib\site-packages\spandrel, replacing 1 existing file. This patch is only tested on the latest chaiNNer nightly, it may not work the standard chaiNNer release.

2.0.0

• V2 release of IllustrationJaNai models for color images includes the following models. The V2 models improve contrast, colors, line smoothness, detail reconstruction, and chroma reconstruction compared to V1:
  • 4x_IllustrationJaNai_V2standard_DAT2_27k (highest quality)
  • 4x_IllustrationJaNai_V2standard_FDAT_XL_18k (balanced quality and speed)
  • 4x_IllustrationJaNai_V2standard_FDAT_M_52k (fastest inference among 4x models)
  • 2x_IllustrationJaNai_V2standard_FDAT_M_unshuffle_40k (2x model with fastest inference)
• DeJPEG Variant: The standard IllustrationJaNai models are trained to denoise halftone artifacts, which may cause unwanted denoise in some instances. These 1x and 2x variants do not denoise halftone artifacts, and they are designed to be as faithful to the original input image as possible, only removing compression artifacts.
  • 1x_IllustrationJaNai-DeJPEG_FDAT_M_100k
  • 2x_IllustrationJaNai-DeJPEG_FDAT_M_100k

The FDAT architecture is not currently supported in chaiNNer, so any models with FDAT in the name do not work in chaiNNer. Use one of the following options:

• Upscale using the latest release of MangaJaNaiConverterGui, which supports FDAT.
• Patch chaiNNer nightly to support FDAT, download spandrel-fdat.zip and extract it to %APPDATA%\chaiNNer\python\python\Lib\site-packages\spandrel, replacing 1 existing file. This patch only works on the latest chaiNNer nightly, it does not work the standard chaiNNer release.

1.0.0

• V1 release of MangaJaNai models for black and white manga pages includes the following models. The model name indicates the height of the input image that model is most optimized to be used on, for example 4x_MangaJaNai_1600p_V1_ESRGAN_70k is best used on manga pages close to 1600px tall. The 4x models are higher quality than their 2x counterparts but run much slower:
  • 2x_MangaJaNai_1200p_V1_ESRGAN_70k.pth
  • 2x_MangaJaNai_1300p_V1_ESRGAN_75k.pth
  • 2x_MangaJaNai_1400p_V1_ESRGAN_70k.pth
  • 2x_MangaJaNai_1500p_V1_ESRGAN_90k.pth
  • 2x_MangaJaNai_1600p_V1_ESRGAN_90k.pth
  • 2x_MangaJaNai_1920p_V1_ESRGAN_70k.pth
  • 2x_MangaJaNai_2048p_V1_ESRGAN_95k.pth
  • 4x_MangaJaNai_1200p_V1_ESRGAN_70k.pth
  • 4x_MangaJaNai_1300p_V1_ESRGAN_75k.pth
  • 4x_MangaJaNai_1400p_V1_ESRGAN_105k.pth
  • 4x_MangaJaNai_1500p_V1_ESRGAN_105k.pth
  • 4x_MangaJaNai_1600p_V1_ESRGAN_70k.pth
  • 4x_MangaJaNai_1920p_V1_ESRGAN_105k.pth
  • 4x_MangaJaNai_2048p_V1_ESRGAN_70k.pth
• V1 release of IllustrationJaNai models for color images includes the following models:
  • 4x_IllustrationJaNai_V1_ESRGAN_135k.pth - good balance of speed and quality.
  • 4x_IllustrationJaNai_V1_DAT2_190k.pth - highest quality model but much slower than the ESRGAN version.

Use MangaJaNaiConverterGui to automatically select the most suitable model when upscaling manga.

Thanks to everyone who made this release possible. Please see the full list of acknowledgements on the main page.

0.0.1

Releasing preliminary upscale model files. These may contain flaws and may not work well on every type of manga image.

NOTE: Newer prerelease models are available on the MangaJaNai Discord server.