This is a LoRA for Ao Kasane from Magia Record. It was trained on the app's character sprites and is designed to be used alongside my Magia Record anime-style LoRA to achieve an anime look. Please note that using the Ao Kasane LoRA on its own will not produce an anime style. The optimal settings are a weight of 0.8 for the Magia Record anime style LoRA and 0.7 for the Ao Kasane LoRA.

Since the standard "open mouth" tag results in a small mouth, using "wide open mouth" is more appropriate. Because the model was trained entirely on open-mouth images, it struggles to render closed-mouth expressions effectively.

LoRA weight is crucial for achieving the anime look. Depending on the settings, you might get an anime style overall, but the hair often retains the coloring style of the app's sprites. To ensure a consistent anime look, the optimal weights are 0.8 for the Magia Record anime style LoRA and 0.7 for the Ao Kasane LoRA. Even without lowering the weights, combining it with the Magia Record anime-style LoRA inevitably leads to some loss in outfit accuracy.

The training images for the Magia Record anime-style LoRA are 1672x957, so please generate images at a large size of 1536x1536 or higher. Generating at smaller sizes, such as 1024x1024, will result in reduced image quality.

Below are the training settings and details.

Since it was trained on the app's sprites, there is no need to convert the art style using Gemini (as was done for Karin Misono), though the facial quality is higher if you do use Gemini for style conversion. As mentioned earlier, the hair tends to revert to the app's coloring style, so it is likely best to manually edit the hair to apply anime-style coloring.

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.

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!

Some magical girl Lora will be hidden if you have if your display settings are set to R/Z/ZZZ. See the list here.