Moonlight Lady - All Characters & Style - Kao no nai Tsuki -Matsuyoi no Soutsubaki-

This is a series of LoRA for the 2025 version of Moonlight Lady (Kao no nai Tsuki -Matsuyoi no Soutsubaki-). See below for details on each of the LoRA on this page.

General Notes

All V1 individual LoRAs are semi-pruned. What does this mean? It means that most of the concepts have been trigger words for the character and outfit, but some have been intentionally left so that they can be edited. Unless I am releasing a LoRA of multiple characters I leave hair style, hair color, breast size, and eye color to where they can be changed. There is also some flexibility built into the outfits as well, but not so much as a completely unpruned dataset would provide.

Note - some of the individual versions have additional notes. Major notes for each release are below, however.

Character Pack

The character pack version is an all-in-one megapack that features the below characters in one LoRA along with functioning as a minor Style LoRA for replicating Carnelian's style. Characters are almost fully pruned so there is less flexibility in the LoRA for augmenting the characters but higher accuracy when prompting multiple characters together. The images used in the showcase are generated with TXT2IMG with ADetailer and HiResFix enabled. I chose similar/identical prompts to the individual LoRA versions to showcase how well the megapack captures the individual characters.

I did not do any cleaning up of images or inpainting for the showcase images. Most images were from the first run of a batch of 3, so anything in the showcase should be similar to what you can expect. The only exception to this is the multicharacter images. I did not clean them up, but I did run a higher batch (8) to select an example for the showcase to get a sample that has less bleed of character concepts across the subjects.

A weight of 1.0 works well enough for most applications, but you can get bad hands/fingers at this weight unless your image is zoomed in to the character. If it is a cowboy shot or full body then the chances of bad fingers/hands is much higher. Reduce the weight, inpaint, or prompt for closer (zoomed in) images to avoid.

Note that not all character combinations will work well. Particularly similar hair colors and other features will cause bleeding. Even those with distinct features will be mixed and matched in the output unless you use additional controls.

Kuraki Suzuna v2

V2 is currently in training. Afterwards I will test it to see how it compares to V1. Either this or Tomomi V2 will be the last individual LoRA before the character megapack which will feature all of the individual characters released + Azuma Io and Hayama. The only possible exception is if v2 fails, as it is a steppingstone dataset towards the megapack.

Major changes with v2 include:

• Slightly reduced file size compared to Suzuna v1

• Addition of Mizuna, Io, and Hayama images in dataset

  • Note Mizuna is very similar to Suzuna. Almost no recognizable difference between the two. Mizuna has slightly softer features and her own kimono outfit. Since they are twins the LoRA does not really pick up on the subtle differences considering all the concepts trained.

  • V1 had a tendency to default any males in images to have a design like Hayama due to overwhelming amount of images with him in it for male subjects. Adding some images of just him as the focus and of Io was done to attempt to allow prompting of the two characters but primarily to add variation in male output for generation.

  • Need more images of just Hayama for the LoRA to train him. This version does not do this well.

  • For Io, use hoodie to prompt his outfit from the remake.

• Additional outfits from v1 (priestess, mizuna's kimono, and pajamas)

  • Priestess outfit is very unreliable. Use see-through to assist prompting. Pruning it did not go well.

• Added some outfits from other characters into the dataset as a test. Ultimate goal is to add outfit control for all characters. Small subset of dataset for v2, so this may not work until a future version. TBD during testing.

Note - I used a new dataset tagging software for this one as CivitAI tagging has been down for several days for me. I went through the results during tag refinement and it seemed to work well. I noticed the facial expressions may not have been tagged as well as I can usually get from CivitAI's tagging system however. Made some manual edits, but it is possible that facial expression control in V1 may be slightly better.

Kuraki Suzuna

I recommend adding the clothing description tags when prompting the outfits. At least include the major pieces like "red skirt". You can prompt other characters in the outfits if you reduce the weight of the character tag and use a LoRA weight of 1.0. This was not primarily intended as an outfit LoRA, so you will see some alteration of the image.

Harukawa Tomomi

Includes a few of her outfits from the game. See trigger words in the version details for trigger words and suggested prompts to trigger them.

Kurihara Sayaka

V1 is a semi-pruned character LoRA for Sayaka. Includes her outfits from the remake: maid, "casual", cheerleader, and ritual/miko attire. Version details has triggers for the character and outfits. Testing with WaiNSFW worked well at weight of 1.0. I was experimenting the training parameters on this one as well as Yuriko's LoRA, so there is a reduced file size.

Kuraki Yuriko

From testing so far weight will need to be pulled back (under 1.0). Weight of 0.80 seems to work fine for prompting the character and her outfit. Definite image distortion as you approach or exceed a weight of 1.0. I was experimenting the training parameters on this one as well as Sayaka's LoRA, so there is a reduced file size.

Sawaguchi Chikako

Seemed to work fine with a weight of 1.0. If you notice distortion of fingers or body parts just pull back slightly to 0.8 ~ 0.9 and it will help in most cases. Trained on everything from the game remake, but predominantly her crop top outfit so that is what I have placed in the trigger words section.