关于Anima：使用tag时，建议追加空格，如[1girl, solo, ]
Memory将会是长期维护的项目，在算力充足的情况下，并且，它完全使用我组合的nai3风格图进行训练，不包含任何画师的作品（3a及之前版本）。
About Anima: When using tags, it is recommended to add a space after commas, e.g. [1girl, solo, ]
Memory is a long-term maintenance project, provided that sufficient computing power is available. It is trained entirely on my curated NAI3-style images and does not include any artist's works (up to and including version 3a).
————————————————————————————————————————
建议的分辨率：832×1344/1024×1536，封面与例图均来自本地部署。
在此感谢我的匿名好友提供的帮助——青秋（Novelai生成工具开发者）。
Recommended resolutions: 832×1344 / 1024×1536. The cover and example images are all from local deployment.
Special thanks to my anonymous friend for the help provided — Qingqiu (developer of the NovelAI generation tool).
————————————————————————————————————————
历史迭代：
同步例图需要将[Clip Skip]设置为1，并启用[A1111]解析(Noob/Illu)。
Version History:
To match the example images, set [Clip Skip] to 1 and enable [A1111] parsing (Noob/Illu).
[lgnis4.5/antwort4]
lgnis版本基于我的另一个项目Elice(antwort4)。
理论上偏向简约线条的同时，也具备更好的构图与泛化能力。
The lgnis version is based on my other project, Elice (antwort4).
In theory, it leans towards simple line art while also offering better composition and generalization capabilities.
[4a.404.h5]：
该版本调整了基础风格的观感，生成结果存在较大变化。
改善了手指与脚趾以及整体肢体的表现，同时在一定程度上增强了提示词的准确性，改善了泛化结果的质量，背景构图等。
This version adjusted the look and feel of the base style, resulting in noticeable changes in the output.
Improved the rendering of fingers, toes, and overall limbs. It also enhanced prompt accuracy to some extent, and improved the quality of generalization results, background composition, etc.
[3a.404.20.65rc27]：
在新的底模上进行了训练，增加了部分游戏角色的标签，整体使用的素材量增加了两倍，
调整了训练素材的比重，因此更偏向可爱风格，相比2a版本也拥有更好的泛化能力。
Trained on a new base model, with additional tags for some game characters. The total amount of training data used has been doubled,
and the proportion of training materials was adjusted, making it lean more towards a cute style. Compared to version 2a, it also has better generalization ability.
[2a.404.02.65/rc]:
使用a100训练了约150小时，它相比1a版本理论上拥有更好以及更稳定的生成结果，兼容之前以及Illustrious与noob的lora模型。
您可以正常的按照Illustrious或noob的使用方式来使用。
Trained on an A100 for about 150 hours. Theoretically, it delivers better and more stable generation results compared to version 1a, and is compatible with previous as well as Illustrious and Noob LoRA models.
You can use it in the same way you would normally use Illustrious or Noob models.
[Lacrimosa Branch]:
非Memory的正确迭代，属于训练过程中出现的一些意外产物，或者说是正式迭代的预览版本。
Not an official iteration of Memory; rather, these are some accidental byproducts that emerged during training, or preview versions of official iterations.
[1a.403.02_3.75]：
基于10w张nai3图片，使用a100进行90小时训练的结果，由于个人算力问题，并未完成最终调优阶段的训练。
Trained on 100,000 NAI3 images using an A100 for 90 hours. Due to limited personal computing resources, the final fine-tuning stage of training was not completed.