This is Locon's style from the Magia Record anime.

When using with the character Lora, it is recommended to lower the intensity to 0.8 or below. High intensity can easily cause distortion or reduce reproduction quality.

The sample images are 1536x1536. Due to their high image quality and resolution, they may appear poor when reduced, but click on them to see their beauty.

CHATGPT told me about the optimal settings and captions. If you're unsure, ask CHATGPT.

Using this Locon style, you can animate illustrations from the Magia Record app. See the bottom of the page for details.

Please note:

Please generate images at a larger size, such as 1536x1536. Generating images at a smaller size, such as 1024x1024, will result in a dramatic drop in quality. This is because the training images are too large, at 1672x952.

Since captions are designed to remove all distinctive features of the character and background, and are intended to be animated, twin tails or two-sided updos may be automatically generated even if you specify a hairstyle. Enter hairstyles you don't want generated in the negative prompt.

The prompt "ocaen" is a typo, but entering it as is will result in a beautiful natural landscape.

You can intentionally bring out the training character's characteristics.

Using the same eye color, hair color, and hairstyle as the training character will help accurately reflect that character's appearance, but this does not work for some characters. Most training images are Iroha, Ui, Mitama, Felicia, and Rena.

Highly recommended:

Similar to Mitama

blue eyes, silver hair, long hair or ponytail

Outstanding cute!Try adding twin tails to your long hair!

Next recommendation

Similar to Felicia

purple eyes, blonde hair, twintails

Similar to Rena

Be careful, a sharp face and angry look makes you look like Rena, and a smile makes you look like Sana.

blue eyes, light blue hair, twintails

Long hair is similar to Iroha, medium hair is similar to Ui

pink eyes, pink hair, long hair or medium hair 

Green hair makes you look like Sana, yellow-green hair makes you look like Momoko, brown hair and long hair makes you look like Touka, red hair makes you look like Kaede, and a side tail makes you look like Kaede and Tsuruno.

If you don't specify eye color, hair color, or hairstyle, there's a high chance you'll get pink or blonde hair with long or medium hair.

Below are the characters used for training

〇 = Likely to appear

△ = Appears, but mixed with characters with the same characteristics

× = Does not appear

△ Akino Kaede

× Alina Gray

× Azusa Mifuyu

〇 Futaba Sana

× Hiiragi Nemu

× Kuroe (1 image)

× Little Kyubey (2 images)

〇 Minami Rena

〇 Mitsuki Felicia

× Nanami Yachiyo (2 images)

△ Satoumi Touka

〇 Tamaki Iroha

〇 Tamaki UI

△ Togame Momoko

〇 Yakumo Mitama

△ Yui Tsuruno

How do I animate Magia Record illustrations?

There are certain conditions: This is intended to be used in conjunction with the Magia Record anime-style Locon. the generated size is large at 1536x1536, and you need to use a Lora trained with "mahou shoujo madoka magica \(anime\)" in the caption of the Magia Record illustration. Also use "mahou shoujo madoka magica \(anime\)" in the generation prompt.

It will look like the animated illustration of Kasane Ao at the bottom of the page. The hair will remain fine, though.

The work tag in the caption is required. If you omit "\anime\" or train it with a different work tag, it will not become an animation. It will become an animation even without a work tag in the caption, but the reproduction quality will drop dramatically. Using Lora alone will not become an animation. Be sure to follow the conditions. Lora posted by other people does not include a work tag, so it will not be able to be animated (I tried it myself and there was no change).

When learning magical girls, we recommend using Lora and lowering the learning rate. For example, Unet 1e-4, Text 4e-6.

Dim must be 32. If it's below 16, the hair will look like the game illustration when animated, and the accuracy will be reduced.

A high learning rate will result in an unanimated result, or the body and face will be animated, with only the hair resembling the game illustration. Learning on a white background will result in the background appearing blurred or weak, so use a gray background or a random solid color.

Settings for 25 Kasane AO ​​images

network_type: LoRA

dim: 32

alpha: 16

unet_lr: 1e-4

text_encoder_lr: 4e-6

num_repeats: 8

epochs: 6

train_batch_size: 2

lr_scheduler: cosine

lr_warmup_ratio: 0.05

min_snr_gamma: 5

What happens when learning with Locon?

The facial expression is strong, but the shadows are occasionally blurred. Furthermore, the costume reproduction quality is lower than with Lora. According to the AI, this is due to structural differences, with the strong influence of Locon, the style of the Magia Record anime, on Lora, resulting in the anime-like result. Additionally, Locon has the tendency to strongly fix shape information. Using the Lora style may result in an unanimated result or reduced reproduction quality.

If you don't understand or fail, ask CHATGPT!