This is Akimabayu from Puella Magi Madoka Magica. There are six types: magical girl, school uniform, knit cardigan, casual clothes, maid, and lab coat.<hr />Added 12/31/2025When used in conjunction with the Magia Record anime style Locon, it becomes even more anime-like. If you use it together, you may not need to read long sentences. It's more anime-like with the dedicated prompts. It's possible to use it without the tedious dedicated prompts, but the shadows, especially around the thighs, will be blurred. This is due to the training images for the style Locon. When using it together, use a larger size, such as 1536x1536, otherwise the image will break down or the accuracy will be reduced. I plan to remake Mabayu next time.Addendum ends here.<hr />Since the training images are game illustrations, this time I used WAI V14 as the base model to resemble the anime. As a result, the design is quite unique. I consulted CHATGPT and optimized it, but this is the limit.This time, I'm making the dataset publicly available, so if anyone has a better experience recreating the Puella Magi Madoka Magica anime than I have with Locon, please let me know. Important points are listed at the bottom.Warning!To recreate the anime illustrations, be sure to follow the instructions below!If you don't, the shading will look like the game illustrations.If this seems long, just follow the sample and you'll be fine!The positive and negative prompts presented below should be used together. Changing the content will prevent it from looking like an anime.Positive prompt<pre><code>((flat 2D anime, cel shading, minimal shading, TV anime style)), soft lighting, even lighting</code></pre>If the prompt in (()) is omitted, it will become a game illustration. Never include soft lighting or even lighting in (()). The face will look a bit strange and will resemble the game illustration.Negative prompts<pre><code>Realistic, 3D, photo, messy shading, extra limbs, poor anatomy, blurry, poor hands, malformed hands, extra fingers, missing fingers, fused fingers, harsh shadow, dramatic lighting, strong contrast, realistic lighting</code></pre>When using a background, it may be a good idea to also add white background, simple background, blurry background, or bokeh. Using these on a white background can cause the shading to lose its animation quality, so do not use them on a white background.To address the high probability of shadows on the body or face when generating the upper bodyThe positive prompts soft lighting and even lighting and the negative prompts harsh shadow, dramatic lighting, strong contrast, and realistic lighting prevent shadows from falling on the face or body when generating the upper body, or prevent silhouettes from appearing on a white background. Use them together.According to CHTPGPT, this is because there are problems with the training images.Step count: 20 or less, CFG 4-5.5, Intensity 0.7 or higherIncreasing the number of steps increases the amount of writing and reduces the animation quality. Increasing the CFG can make the face and painting look unnatural. If you don't mind a lower level of fidelity, a level below 0.6 is acceptable.Never omit the trigger word "mahou shoujo madoka magica"!It's included in the caption to approximate the style of Puella Magi Madoka Magica, so omitting it will significantly reduce the accuracy of reproduction.Don't use the quality tag!Using the quality tag will cause the AI ​​to try to increase the image quality, which will result in the image no longer looking anime-like. It should also not be used for negative prompts.Other detailsSince the training is done on a white background, the background may be weak, simple, or the face may appear subtle. If the face appears subtle, add "detailed face" and "sharp focus" after the positive prompt. Never put them in parentheses (()) because they will lose their anime-like quality. Do not use them on a white background, as they will make the face appear subtle.They look the most anime-like when dressed in gothic lolita or other clothing that is outside the scope of the study. Explaining why this is the case would require several times more text, so if you have any questions, just ask CHATGPT! You'll get a satisfactory answer!Use clips as desired. Setting 1 may be too strong, while setting 2 may deviate from the anime. Depending on the prompt combination, the face may appear unrecognizable or the fidelity may decrease. The knit cardigan has a high degree of reproduction when worn alone, but when worn in multiple costumes, the cream-colored lines are not reproduced as well, so it is rated (knit cardigan: 1.1).Notes for CreationLocon is recommended when creating multiple outfits.For Locon, the following settings are recommended:XL Lora Trainer by HollowstrawberryDefault settings are fine unless otherwise specified.Base model: WAI V14Repeat count: 6Epochs: 12 (Using Epoch 10. 11 or more is overfitting. 8 is just the appearance, but has poor reproducibility when lowering the intensity. 9 is a safeguard.)Unet: 3e-4Text encoder: 3e-5batch size: 4Lr scheduler: Cosine with restartsLr scheduler number: 3Lr warmup ratio: 0.03dim: 16 alpha8 conv dim: 16 conv alpha8When creating with Lora, the knit cardigan's reproduction is low and the cream-colored lines are barely reproduced (it is reproduced when used alone), and the green hair accentuates the green features. Even with Locon, reproducibility was low, and I struggled to find optimal captions.A dim of 8 results in a subtle face and reproduction quality, so dim 16 Alpha 8 is recommended. According to CHATGPT, Alpha 16, dim 32 or higher, and Lycoris are not recommended because they overfit and lose their anime-like quality.Again, if you create multiple outfits with Lora, the knit cardigan will definitely be a green knit cardigan and the cream lines will not be reproduced.The captions are based on Locon, so creating multiple outfits with Lora will result in a lower reproduction of the maid outfit (you'll need to change the caption). When creating with Lora, a frilled brown striped dress is not acceptable; you need to separate the brown maid outfit, striped outfit, and brown striped skirt. Otherwise, the outfit and stripes will not reproduce well.It seems that frilled is best worn behind a white waist apron.Adding frilled to clothes or skirts would overlap the meaning, so it's a no-no. This is due to the structural differences between Lora and Locon. All characters have been Locon since Seitokai no Ichizon. According to CHATGPT, SD1.5 is more likely to be anime-like than the SDXL series.When creating other Magia Record characters or Magia Record-exclusive costumes, special designs like this will be required. Excluding the Magia Record anime version.

Files with SHA256: a267d29da7e9b92e2216f1fd1596ba408d0aa94d55287f1b3ede49cb4e594a33