Neon (NIKKE) LoRA

This is a character LoRA model for Neon from Goddess of Victory: NIKKE. It is trained against the original SD 1.5 checkpoint, so it should be compatible with many base models derived from SD 1.5. This version, 0.2, is still a snapshot from an incomplete work. It works in some degree but not fully.

When compared to its version 0.1, this version:

• added more sample images while removed some less suitable ones so that it can infer more varieties of images,

• modified some sample images, hoping it infers Neon more consistently,

• fully rewrote captions for the samples so that it can react prompt better,

  • in particular, captioned Neon herself, her costume, and her weapon separately,

• balanced the arts and (cosplay) photos in the samples so that is better fits to both types of inferences,

• and analyzed the learning stats more carefully to get it to the optimal learning steps.

You can now infer Neon more similar to her appearances in the game easily, and you will need less description words for Neon in your prompts, e.g., sailor collar or mini skirt. (Although you may still need some.) You will not need to specify neon lights in your negative prompt either.

This version has three separate trigger words: neon (nikke) for Neon herself, neon (nikke) costume for her default outfit, and neon (nikke) shotgun for her default weapon (the one similar to SM DP-12 and called Annihilator in English and 올 브레이크 (All Break) in Korean, though this model doesn't know these names). If you use automatic1111 or compatible prompt syntax, you need to escape the parentheses by backslashes, so you need to write them like neon \(nikke\), neon \(nikke\) costume, and neon \(nikke\) shotgun. However, you can often abbreviate them to just neon, neon costume and shotgun for this LoRA to take effect. (The magical power of CLIP, perhaps.) If you want a cosplay photo, adding cosplay in your prompt will help.

Suggested strength of this LoRA is around 0.8.

See the generation info of example images to know my intended prompt patterns. I captioned with both full English sentences and danbooru tag sequences, so (hopefully) either style of prompts works. Note that some example images are inferred using ControlNet (though Civitai doesn't show you ControlNet-related parameters in the generation info.)

Some cautions:

• This model is very bad to produce images of Neon's shotgun. Although slightly better than v0.1, v0.2 frequently infers very exotic object that is only slightly similar to Neon's shotgun...

• This model doesn't know Neon's alternative outfits in the game. In particular, the training set contains no images for Bling Bullet or Blue Ocean versions of Neon. (Sorry for swimwear funs!)

• Although I have no will to restrict uses of this LoRA model in any way, the Neon character appears an intellectual property of the right holder(s) of the game. There is a legal document called Secondary Creative Activities Guideline (https://policy.shiftup.co.kr/ip/en/index.html (in English) and https://policy.shiftup.co.kr/ip/kr/index.html (in Korean)), which is published by the company that created the game and discusses the use of the characters from Goddess of Victory: NIKKE. You should be aware of it when you use my Neon LoRA model.

Have fun!

Neon: Master, why is nobody creating my LoRA model? Most of the Nikke LoRA are for SSR characters, and I know they deserve it. But Rapi has two, and Anis has three, while they are SR like me! Because they are members of Counters, the "protagonist" squad? Then, I am, too! What's wrong with me? Because I'm a spy? Because my firepower is not enough?

Commander: Calm down, Neon. Actually, I'm doing it, though I need some more time before finishing it. Also, you will soon be the first in the squad to have an SSR variant. Do you know it?

Neon: !!

Noen: That's great. Thank you, Commander!

Commander: (Well, "which" is great for Neon more?)

Putting nonsense aside,

This is a LoRA model for Neon from Goddess of Victory: NIKKE. It is trained against the original SD 1.5 checkpoint, so it should be compatible with most of the base models derived from SD 1.5. Please note, however, that this is just a snapshot from an incomplete task. It works in some degree, but not fully.

The trigger word is "neon (nikke)", and you need to specify it "neon \(nikke\)" if you use automatic1111 or compatible prompt syntax. Adding some descriptive words helps inferring an image similar to her default visuals in the game. Examples are: glasses, beret, blue bow, brown choker, white shirt, sailor collar, blue collar, blue tie, blue kerchief, khaki gloves, white skirt, miniskirt, belt pouch, khaki pouch, ammunition pouch, thighhighs, high heels, white footwear. Also, it may be a good idea to specify "neon lights" in the negative prompt, unless you want a neon-lit cyberpunk scene, for example.

Suggested strength is 0.6-0.9, and Clip skip can be 1 or 2, depending on the base model and other factors. It seems that applying to a photorealistic model requires more strength (sometimes more than 1.0).

Any suggestions or other feedback is welcome, especially if Neon is your fave (like me. :)

Have fun!