底模：Pony
主页：https://www.liblib.art/modelinfo/569b499c49294194857e80c75026301a
版本：v2.0
Collection of All Character in One Models 全角色模型合集
v2.0版本使用5w张图片，在非常充足的算力--4*4090下训练了超过150个GPU小时。
触发词
arknights
all in one
v2版本支持的角色大为提高，包含240个有效的角色文件夹，经评估，支持150~200位明日方舟的有效角色。
v2版本在v1.5版本基础经过多次加练（这也是为什么文件名叫v4）
该模型的训练耗时半个月，花费了大量的人力和物力，将支持角色从80位拓展到150~200位。
如果有机会，该模型依然会不断更新，
特别鸣谢/ 算力兼技术支持：昊天
非常感谢 昊天 同学对本项目的大力支持！
角色数据集使用 waifuc工具搜集。
数据集的标签排序使用 waifuset工具完成。
角色后期的步数平衡在的昊天帮助下完成。
CCIP（角色的还原度）的测试结果：CCIP Test Results
提示词文档：明日方舟全角色使用指南 Guide for All Characters in Arknights | Civitai
Version 2.0 used 50,000 images and was trained for over 150 GPU hours on a powerful hardware setup - 4*4090.
Version 2.0 greatly expanded the supported characters, with 240 valid character folders. Based on evaluation, it supports 150 to 200 valid Arknights characters.
Version 2.0 underwent multiple iterations on top of version 1.5 (which is why the file name is v4).
The training of this model took half a month and required a significant amount of manpower and resources to expand the character support from 80 to 150 to 200.
If given the opportunity, this model will continue to be updated.
Special thanks to Hao Tian for their support in terms of computing power and technical assistance.
A big thank you to fellow student, Hao Tian, for their strong support in this project!
The character dataset was collected using the waifuc tool.
The sorting of labels in the dataset was done using the waifuset tool.
The balancing of post-processing steps for character animations was done with the help of Hao Tian.
CCIP (Character Completion Index Percentage) test results can be found here: CCIP Test Results
Guide for All Characters in Arknights: Guide for All Characters in Arknights
v1.5版本使用20K图片，在v1.0的版本基础上，使用双卡A100 80G 训练了35个GPU小时。
1.5版本共优化了两位角色，新增了167位角色。（85+167）
新增的167位角色有50%~70%的还原度，但不能完全还原，可能是因为训练策略的问题。
更多角色，请期待2.0版本！
其中，有70位角色具有100张图像，每张图被训练了20*2次
有87位角色具有50张图像，每张图被训练了40*2次
羽毛笔和蓝毒的默认服装在v2.0版本中得到优化。
提示词文档：明日方舟全角色使用指南 Guide for All Characters in Arknights | Civitai
该模型永久免费下载。
获得作者许可后可以转载模型，但不得：
冒充作者、付费下载、使用模型盈利，以及一些违反法律法规或做出不当的行为。
Version 2.0 utilized 20,000 images and was trained on dual A100 80GB GPUs for a total of 35 GPU hours, building upon the foundation of version 1.0.
Version 2.0 included optimization for two existing characters and introduced 167 new characters, resulting in a total of 85 + 167 = 252 characters.
The newly added 167 characters achieved a restoration level of 50% to 70%, but it is not a complete restoration. This may be due to the training strategy employed.
More characters to come! Stay tuned for version 2.0!
Among them, 70 characters had 100 images each, with each image being trained 20 * 2 times.
87 characters had 50 images each, with each image being trained 40 * 2 times.
The default costumes of Feather and Blue Poison from Arknights received optimization in version 2.0
The model is available for permanent free download.
With the author's permission, the model can be reproduced, but certain actions are prohibited:
-
Impersonating the author
-
Charging for downloads
-
Using the model for profit-making purposes
-
Engaging in activities that violate laws and regulations or inappropriate behavior
v1.0版本使用超过3w张图片，在A100 40G 上训练了超过80个GPU小时。
v1.0版本共支持85位角色，后续会持续新增角色，直到达到预期值--238位角色。
v1.0版本的85位角色，是经过我的数据库统计，单人图片多于400的85位角色。
部分角色在默认皮肤上表现欠佳，以我目前的能力，无法将默认皮肤单独筛选出来。
全角色项目长期缺人手，如果您有兴趣，请联系我。
“让每个知名系列的角色都有他的lora，让每一份热爱都可以实现”，
这个是全角色项目的最开始的出发点。
The v1.0 version was trained using over 30,000 images and consumed over 80 GPU hours on an A100 40G.
The v1.0 version supports a total of 85 characters, and more characters will be continuously added until the target of 238 characters is reached.
The selection of the 85 characters in the v1.0 version was based on my database statistics, focusing on characters with over 400 individual images.
Some characters may not perform well with their default skins, and currently, I am unable to filter out the default skins separately based on my current capabilities.
The project for all characters is in need of long-term assistance. If you are interested, please contact me.
The initial goal of the all characters project is to "give every well-known series its own Lora and make every passion come true."
Since I am working on this project in my spare time, I cannot create individual example images for each character. Therefore, I rely on feedback and optimization suggestions from users for the model's improvement.
Please note that this model's content includes some covers that have been rated as PG-13/R/X
Please choose the appropriate age rating based on your needs.
Training model: Pony
Recommend model: Pony Diffusion V6 XL/T-ponynai3/AutismMix SDXL
Training resolution: 1024*1024
Recommended output resolution: 832*1216
特别鸣谢/算力 : Nyan9的显卡租借/我的钱包
Special Thanks/Computational Power: GPU Rental/My Wallet
Discord : https://discord.com/invite/CFpyHceF52
Recommended dataset collection and processing tools: https://github.com/deepghs/waifuc
Possibly the most user-friendly dataset processing UI currently in use:https://github.com/Eugeoter/waifuset