This is a model related to the "Umamusume" characters, but I'm not aiming to replicate a specific character. Instead, I'm trying to simulate the in-game 3D models. However, due to the limited dataset, the model can still clearly resemble certain characters. Since I didn't label these in the dataset, the character generation is actually quite random. I'll improve these issues in future versions.
これは「ウマ娘」キャラクターに関連するモデルですが、特定のキャラクターを再現するのが目的ではなく、ゲーム内の3Dモデルをシミュレートすることを目指しています,しかし、データセットが少ないため、モデルは依然として特定のキャラクターをはっきりと再現できる場合があります,データセットでこれらにラベル付けをしていないため、実際にはキャラクター生成がかなりランダムになっています,今後のバージョンでこれらの問題を改善していきます,
这是一个关于“赛马娘”角色的模型，但是我并不是去为了还原某个角色，而是为了模拟游戏内的3D模型，但是由于数据集较少，实际上模型还是可以明显的还原出某些角色，但是由于我并没有在数据集中对这些进行打标，实际上角色生成时非常具有随机性，以后得版本里我会改进这些问题,

This is just a test version. I didn't rigorously label the dataset, and the annotation prompts I used were quite casual. Even so, it has achieved decent results, so I decided to share it. If you're interested, feel free to give it a try. Below are some recommended parameters and points to keep in mind.
これは単なるテストバージョンです,データセットのラベル付けは厳密に行わず、使用した注釈プロンプトもかなりカジュアルなものでした,それでも、なかなか良い結果を出したので、共有することにしました,もし興味があれば、ぜひ試してみてください,以下に、推奨パラメータと注意点をいくつか挙げます,
这仅仅是一个测试版本，我并没有十分严谨的进行数据集打标，使用的标注提示词也十分随意，但是他仍然取得了不错的效果，因此我把他分享出来，如果你感兴趣可以试一试，下面是一些推荐的参数和需要注意的地方,

There are no specific requirements for "Step" and "CFG." You can experiment freely based on your own experience.

sampler : eular A(When testing other samplers, I often encountered unexpected results. Therefore, I recommend sticking to the ones I suggest.)

Regarding image sizes, you might notice that the face occupies a relatively small portion of the generated image. Additionally, other factors can lead to smaller-sized images not achieving optimal results.

First, if you want to create images similar to the example ones, ensure that the aspect ratio of the image is close to 3:5. Second, aim for larger image dimensions (or use facial restoration). For instance, in the example images I generated, dimensions like 1088x1772 or 960x1600 were used. With these settings, the generated images generally do not require additional facial retouching, and the character proportions appear more natural.

prompt:

The training set includes four basic types of clothing. As for other complex outfits, similar to different characters, you might be able to generate them successfully by using appropriate prompts. However, for the V1 version, the results you get will be highly random. :)

The four basic types of clothing are shown in the image below.

Here are some examples of prompts. These are just simple examples.

score 9, score 8 up, score 7 up, umamusume,white background, uma costum, 1girl, horse ears, animal ears, horse girl, solo, cabbie hat, hairclip, shorts, hat, hair ornament, brown hair, horse tail, tail, thighhighs, 

score 9, score 8 up, score 7 up, umamusume,white background, uma full dress, 1girl, animal ears, horse ears, solo, horse girl, tail, black hair, horse tail, thighhighs, boots, long hair, full body, hand on hip, looking at viewer, 

I once tried generating non-full-body images to achieve better results with smaller-sized images. I added prompts like "upper body" or "cowboy shot" in the positive prompts and included "full body" in the negative prompts. This has a chance of producing the desired structure, but it's still quite unstable. This is a possible solution approach, and if you're interested, you can give it a try.

Feel free to try this model out, and share your generated results in the comments section. If you have any suggestions, please post them in the comments as well.

描述:

训练词语: umamusume