Shounen Zoom (Shigeru) Anime Style for PonyXL

Now updated to v6i! See updated usage/tag notes below:

This LORA is design to simulate the style of Shounen Zoom by Shigeru. This new version is much better at recreating the style and gives you more control over style elements. No trigger words needed, but read the description of the score tags for details. This LORA was trained using OneTrainer's multi-resolution feature at 1024, 1088, and 1152 resolutions. This should help with image sharpness and details. It is recommended that you use a HiRes fix pass to get the most detail (I suggest Latent Bicubic at ~0.55 denoise).

Recommended to be used at 0.95 strength. It may be a little overbaked, so full strength tends to mess up hands. If using a non-Autism mix checkpoint, you may need to go over 1.0 strength to maintain the style.

Please support the original artist wherever possible so they can continue to produce their art. Without them, LORAs such as these would not be possible.

Now also available for generation on TensorArt: https://tensor.art/models/746261545887177583

USAGE NOTES:

V6i of this LORA was trained specifically on Autism Mix and images were ranked in scores 5-9. Shigeru's multiple styles were tagged with "2020s" "2015s" and "2010s." There was a mistake in tagging where all score_9 "2015s" images also included the "2020s" tag, but this improved the quality of the "2020s" generations, so I'm still posting this. If you want to recreate the middle style of Shigeru, you will need to use both tags.

NOTE: Since it seems you all want better NSFW ability, the new v6i version uses a largely uncensored dataset. I decensored the dataset using the previous versions of the LORA so you should get better NSFW generation without the need to add various censors to the negative (a simple "uncensored" tag in the positive should do the trick).

For v4d: An inpainting pass with "mosaic censoring" or "bar censor" in the negative may help with this. For this first pass, images were not tagged with rating_safe, rating_explicit, etc. Maybe in the next version.

What happened to v5? I injected far too many manga scans and it wrecked the style, but it worked great for inpainting and decensoring the dataset. Because of this, I didn't release it v5, but went straight to v6.

TAGGING NOTES:

v6i tagging:

Scores 9 and 8 are the most recent works by the artist.

Scores 6 and 7 start to add some older works using the "source_watercolor" and "2010s" tags.

Score 5 is almost entirely manga scans that provide some extra detail, but can otherwise mess up the image quality (mostly in terms of shading). These are tagged with "source_manga" and some also have an "ai upscaled" tag. It is recommended to not include a "score_5_up" tag, but it may sometimes help with hands.

V4d tagging:

Scores 9 and 8 are the most recent works by the artist that most reflect his current visual style.

Score 7 is mostly sourced from manga scans that were uprezzed using AI. You can prompt for this using "source_manga" and monochrome

Score 6 is mostly the artists earlier works that may be slightly different in style. Lots of these were tagged using "source_watercolor" to reflect his use of watercolor for coloring compared to the current style.

Please do not violate the content policy -> https://civitai.com/safety#content-policies

Please do not post NSFW images of minors to this model page