ヒトこもる(Hito Komoru) Style LoRA

中文 | English 使用前阅读 | Read it before use

--

• 这个模型对小分辨率的图片生成不是很友善，尤其是full_body的时候，这可能需要最终画面将近1080p来达到比较好的效果，Denoising strength ∈ [0.45, 0.8)。

• This model is not very friendly to generating small resolution images, especially for full_body images. To achieve a better effect, the final image should be close to 1080p, with Denoising strength ∈ [0.45, 0.8).

--

Recommended:

Style1:

Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11.5, Model hash: 1065c5d2e0, Model: CreationMixV1, Denoising strength: 0.45, Clip skip: 2, Hires upscale: 1.6, Hires steps: 15, Hires upscaler: Lanczos

Style2:

Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 9.5, Model hash: 1065c5d2e0, Model: CreationMixV1, Denoising strength: 0.62, Clip skip: 2, Hires upscale: 1.6, Hires steps: 15, Hires upscaler: Latent

--

前言

仿ヒトこもる作画风格的LoRA模型。主要强化对full_body的训练，当然，upper_body也没问题，支持多种纵横比。

Introduction

LoRA model imitating the drawing style of Hitoimim. The primary focus is on training for full_body, but upper_body is also supported, with compatibility for various aspect ratios.

--

关于这个模型的训练

full_body的素材图很少，但是经过之前对米山舞的风格模型训练后，我能更加从容地应对诸如此类的问题了。同以前一样，我把很多词条都拆分了出来，如果你不使用这些词条，那么这些效果就不会显示出来，我认为这个能提高图片生成的可塑造性。

Training of this model

There are very few source images for full body, but after training the model YoneyamaMai style, I can now more confidently tackle such problems. As before, I have separated many entries, and if you do not use these entries, then these effects will not be displayed. I believe this can increase the malleability of image generation.

--

有用的提示

• 当你生成的图里的人物脖子非常长的时候，有这个解决办法，添加如下词条：breast, dress。

• 一些时候你会发现，头发“耷拉”着很难看，这时候也可以尝试添加一个或几个词条改善这种情况，比如 silver_hair, wind, hair_undone, hair_half_undone 等等，这样是不是好多了？此外，还有比较简单的解决方法，那就是使用2.5D的Checkpoint模型。在我看来，画风模型更多的时候提供的是基础，而不是将部分效果直接表现出来。好用的东西往往有其局限性，功能丰富的东西往往不太好上手。

• 如果遇到看不出LoRA画风的情况，请考虑更换Checkpoint模型，尤其是一些自带强烈风格的Checkpoint模型。如果出图相当不稳定，也应该考虑更换Checkpoint模型。

• 我训练的大多数2D LoRA 风格模型都没有经过3D Checkpoint模型的测试。不过如果有强烈的兴趣，也不是不能一试。

• 如果你在使用badhandv4 (Textual Inversion)，并且你生成的图四肢出现畸形，暂时停用它以改善这一情况。一般只有在原图的四肢画的不怎么好的时候才应该考虑使用badhandv4。

• 还有疑问直接在评论区提出就行，不过由于该模型仍然有其局限性，有些问题一时半会儿估计也不太好解决，我尽力而为，能帮上忙就帮帮。

Useful Tips

• When the character's neck in the generated image is excessively long, there is a solution to this. Add the following parameters: breast, dress.

• Sometimes you may find that the hair hanging down doesn't look good. In this case, you can also try adding one or more parameters to improve the situation, such as silver_hair, wind, hair_undone, hair_half_undone, etc. Isn't it much better? In addition, there is a relatively simple solution, which is to use the 2.5D Checkpoint model. In my opinion, the painting style model mainly provides a foundation rather than directly presenting some effects. Things that are easy to use often have limitations, and things with rich functions are often difficult to start with.

• If you encounter cases where the LoRA style isn't evident, consider switching Checkpoint models, especially for those that come with strong styles. If the output is unstable, you should also consider changing the Checkpoint model.

• Most of the 2D LoRA style models I trained haven't been tested with 3D Checkpoint models. However, if you're strongly interested, it's worth a try.

• If you're using badhandv4 (Textual Inversion) and your generated images have deformed limbs, temporarily disable it to improve the situation. Generally, you should only consider using badhandv4 when the limbs in the original image aren't drawn very well.

• Please feel free to ask any questions in the comments. However, as this model still has its limitations, some issues may not be easy to solve immediately. I'll do my best to help if I can.

--

模型主要信息

权重建议值：1 (or 0.5~1.4)

基于模型：Anything-5RE

训练信息：筛选出共计276张原画（含裁剪图），迭代808步，多分辨率、智能定向、集散词条、梯度迭代训练。

其他的一些建议：EasyNegative。此外，万一畸形四肢换个种子重画就行。

About Model

Recommended weight value: 1 (or 0.5-1.4)

Based on model: Anything-5RE

Training info: Selected 276 original works (including cropped images), iterated 808 times, trained with multiple resolutions, intelligent orientation, distributed keywords, and gradient iteration.

Other recommendations: EasyNegative. Also, if you encounter deformed limbs, try with a different seed. It's a painting style model.

--

如果你创作出了好作品，欢迎分享它。如果它无法满足你的需求，我是说，我的朋友，请不要给出一星好评。

If you've created something good, feel free to share it. If it doesn't meet your needs, I mean, my friend, please don't give a one-star 'praise'.