Sign In

Chenkin Anime Hotbaby

525
4.1k
112
Updated: Mar 26, 2024
base modelanimewoman
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
1,621
Reviews
Published
Mar 26, 2024
Base Model
SDXL 1.0
Hash
AutoV2
E73324E3D1
Silver Creator Badge
Chenkin

Chenkin Anime Hotbaby V2: Chenkin Anime Hotbaby V2 is based on Animagine XL 3.1, utilizing hierarchical label training to achieve a more accurate representation of human anatomy and greater flexibility in the anime text-to-image model. The model inherits the advantages of Animagine XL 3.1 in terms of hand anatomy, efficient label sorting, and enhanced anime concept knowledge.

It is also compatible with most Lora models trained on Animagine XL 3.1.

In Chenkin Anime Hotbaby V2, the majority of characters supported by Animagine XL 3.1 can be used.

In Chenkin Anime Hotbaby version, due to the absence of a clear distinction between SFW and NSFW concepts, when users provide simple prompts, it is prone to generating images with revealing or uncovered clothing.

We deeply apologize for this. Therefore, in the V2 version, we are training a model with explicit grading and a healthier approach, allowing users to use it with peace of mind.

Chenkin Anime Hotbaby V2 was trained using approximately 45,000 images and follows a strict grading system, with grading labels given top priority. During training, there are four main grading labels. In simple terms, these four gradings can be understood as:

general	                 R12-  大众级,不暴露的
sensitive	         R15-  广义的大众级,较少暴露的
questionable / nsfw	 R18-  裸露的,比较暴露的
explicit	         R18G- 色情的,非常暴露的

The corresponding image quantities for these four gradings are as follows:

general	                 5K
sensitive	         5k
questionable / nsfw	 17k
explicit	         17k

How to use these grading labels?

For example, when we want to generate an image of Asuka Langley from Neon Genesis Evangelion, we typically use the following prompt:

1girl, solo, souryuu asuka langley, neon genesis evangelion

If we want the character's outfit to be more revealing:

sensitive, 1girl, solo, souryuu asuka langley, neon genesis evangelion

If we want the character to be depicted without covering the clothing:

questionable, 1girl, solo, souryuu asuka langley, neon genesis evangelion

If we want the character to engage in behavior unsuitable for public places (strongly discouraged):

nsfw, explicit, 1girl, solo, souryuu asuka langley, neon genesis evangelion

It is worth mentioning that this model is a wholesome model that you can use in public settings. Just:

general,sfw,

Include the following gradings in negative prompts to avoid generating revealing image

nsfw, sensitive, questionable, explicit

For specific usage of prompts, you can refer to the negative prompts used in Animagine XL 3.1 to guide the model in generating aesthetically pleasing images.

nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

To obtain higher-quality results, add the following content before the prompts:

masterpiece, best quality, very aesthetic, absurdres

Advantages of V2 version compared to V1 version:

1.Better controllability, suitable for a wider audience

With enhanced grading labels, you can freely switch between "sfw" and "nsfw".

2.More characters, supporting over 5000 characters from popular IPs

Refer to the supported character reference list:

Supported Characters

This is made possible because V2 version is trained based on Animagine XL 3.1, which is one of the most advanced anime models currently.

3.Improved anatomical structure

In the V2 version, the model can more easily differentiate between the concepts of "anus" and "vagina" and is less likely to confuse their positions.

4.Richer knowledge in certain domains

In the "explicit" dataset, we have included a diverse range of content, including but not limited to:

tentacle

futa

Anal

Fff_threesome

...

Training configuration

The training parameters of this model follow the configuration of the V1 version.

However, some adjustments have been made in the optimizer_arguments section.

[optimizer_arguments]
optimizer_type = "AdaFactor"
learning_rate = 7.5e-7
train_text_encoder = true
learning_rate_te1 = 3.75e-7
learning_rate_te2 = 3.75e-7
optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False",]
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 100
max_grad_norm = 0

Acknowledgments

Special thanks to GPU Sponsor: nieta.art / Neta

Respected model team: cagliostrolab

Thanks to the predecessors who provided guidance and assistance during the model process:

@narugo1992, whose waifuc was the main tool for collecting and cleaning the training dataset.

@L_A_X , for naming and providing conceptual guidance for the Hotbaby series.

@Euge , whose waifuset was the main tool for data grading.

Acknowledgment to the community: XL Open Paradise of the 2D World

Acknowledgment to the team: NAIXL

Special Reminder

Do not repost or publish this model on any other platform without the author's permission!

The reuploader and corresponding platform are responsible for any negative consequences resulting from this.

2.0版本更新:

Chenkin Anime Hotbaby V2 是基于Animagine XL 3.1训练,采用分级标签训练,追求更准确的人体结构和更高的自由度的动漫文本到图像模型。模型在手部解剖结构、高效的标签排序和增强的动漫概念知识方面,继承了Animagine XL 3.1的优点,同时也兼容大部分基于Animagine XL 3.1训练的Lora模型。在 Chenkin Anime Hotbaby V2 中,可以使用Animagine XL 3.1支持的绝大部分角色。

Chenkin Anime Hotbaby 版本中,因为没有明确区分SFW 和 NSFW 的概念,用户直接使用简单的提示词时,容易生成穿着暴露服装或者不覆盖服装的图片。

对于这一点,我们深感抱歉。

因此,V2 版本我们此在训练一个明确分级、更加健康的模型,让用户安心放心使用。

Chenkin Anime Hotbaby V2 共使用约45k图片进行训练,采取严格的分级制度,并将分级标签放到最优先的位置。

在训练中,共有四个最主要的分级标签。

这四个分级通俗的说,可以理解为:

general	                 R12-  大众级,不暴露的
sensitive	         R15-  广义的大众级,较少暴露的
questionable / nsfw	 R18-  裸露的,比较暴露的
explicit	         R18G- 色情的,非常暴露的

他们与图片数量的对应关系如下:

general	                 5K
sensitive	         5k
questionable / nsfw	 17k
explicit	         17k

如何使用这些分级标签?

举个例子,当我们想生成明日香的图片时,通常使用这样的提示词:

1girl, solo,souryuu asuka langley, neon genesis evangelion

如果我们想让该角色的服装更加清凉:

sensitive,1girl, solo,souryuu asuka langley, neon genesis evangelion

如果我们想让该角色不覆盖衣服:

questionable,1girl, solo,souryuu asuka langley, neon genesis evangelion

如果我们想让该角色发生一些不太适合在公众场所进行的行为(非常不建议):

nsfw,explicit,1girl, solo,souryuu asuka langley, neon genesis evangelion

值得一提的是,该模型是一个健康的模型,您完全可以在公开场所使用,只需:

在正面提示词中加入:

general,sfw,

并在负面提示词中加入如下分级,避免出现暴露的图片:

snfw,sensitive,questionable,explicit

在提示词的具体使用上,可以参考Animagine XL 3.1 使用的负面提示,以引导模型生成高美观的图像

nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

为了获得更高质量的结果,请在提示前添加以下内容:

masterpiece, best quality, very aesthetic, absurdres

V2 版本相比 V1 版本的优点:

1.更好的可控性,适合更多人群

借助经过强化的分级标签,您可以自由地在“sfw”和“nsfw”中自由切换

2.更多的角色,支持各热门IP的5000+位角色

支持的角色参考列表:

https://huggingface.co/spaces/cagliostrolab/animagine-xl-3.1/tree/main/wildcard

这得益于V2 版本基于Animagine XL 3.1 进行训练,他是当前最优秀的动漫模型之一

3.更好的生理结构

在V2 版本版本中,模型已经可以比较容易地区分”anus“ 和”vagina“的概念,并且不容易将他们的位置混淆。

4.某些领域更丰富的知识

在”explicit“数据集中,我们加入了非常多元的内容,包括但不限于:

tentacle

futa

Anal

Fff_threesome

....

训练配置

在训练参数上,此模型沿用了V1版本的训练配置。

但在optimizer_arguments部分,做了少许调整。

[optimizer_arguments]
optimizer_type = "AdaFactor"
learning_rate = 7.5e-7
train_text_encoder = true
learning_rate_te1 = 3.75e-7
learning_rate_te2 = 3.75e-7
optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False",]
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 100
max_grad_norm = 0

鸣谢名单

特别鸣谢:算力赞助 nieta.art / GPU Sponsor: Neta

最尊敬的模型团队:cagliostrolab

感谢在模型过程中提供指导和帮助的前辈:

@narugo1992 他的waifuc 是我搜集和清洗训练集的主要工具

@L_A_X Hotbaby系列命名并提供思想指导

@Euge 他的waifuset是我给数据分级的主要工具

鸣谢社群:二次元的XL开放乐园

鸣谢团队:NAIXL

特别提醒

未经作者许可,请勿将此模型转载或以任何方式发布到其他平台!

由此产生的一切不良后果,由转载者以及对应平台承担。


Overview

Chenkin Anime Hotbaby is an anime text-to-image model based on Animagine XL 3.0, aiming for improved human body structure and higher degrees of freedom. The model inherits the advantages of Animagine XL 3.0 in hand anatomy, efficient label sorting, and enhanced anime concept knowledge. It is also compatible with most models trained on Animagine XL 3.0. In Chenkin Anime Hotbaby, you can utilize the majority of characters supported by Animagine XL 3.0.

Reference Animagine XL 3.0, here is the list of all supported characters: https://huggingface.co/spaces/Linaqruf/animagine-xl/blob/main/wildcard/character.txt

Chenkin Anime Hotbaby is trained on a dataset of 53k high-quality anime images, using over 150 GPU hours on an A40 48G.

Regarding the training dataset, a significant portion consists of nudity, as the author believes this content should be part of a comprehensive model to better understand human body structure.

When not specifying clothing or using character names directly, there is a high probability of generating explicit images. Therefore, it is strongly advised to use this model with caution in public settings.

Training Configuration

For specific usage of prompts, you can refer to the negative prompts used in Animagine XL 3.0 to guide the model in generating aesthetically pleasing images.

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name

To obtain higher-quality results, please include the following content before your prompts::

masterpiece, best quality

The training configuration mostly follows that of Animagine XL 3.0, with slight adjustments in learning rate and training of the text encoder.

Acknowledgments

Special thanks to GPU Sponsor: nieta.art for providing computational resources.

Respectful acknowledgments to the model team: cagliostrolab.

Dataset provided by: Miss Er.

Thanks to the mentors who provided guidance and assistance during the model development:

@kblueleaf

@L_A_X

@Euge

@LEOSAM

Acknowledgments to the community: 二次元的XL开放乐园

Acknowledgments to the team: NAIXL

Special Note

Please do not repost or distribute this model on other platforms without the author's permission!

The reuploader and the corresponding platform will bear all responsibilities for any negative consequences resulting from such actions.

概述

Chenkin Anime Hotbaby是基于Animagine XL 3.0训练,追求更好的人体结构和更高的自由度的动漫文本到图像模型。模型在手部解剖结构、高效的标签排序和增强的动漫概念知识方面,继承了Animagine XL 3.0的优点,同时也兼容大部分基于Animagine XL 3.0训练的模型。在 Chenkin Anime Hotbaby中,可以使用Animagine XL 3.0支持的大部分角色。

参考Animagine XL 3.0,支持的所有角色列表:https://huggingface.co/spaces/Linaqruf/animagine-xl/blob/main/wildcard/character.txt

Chenkin Anime Hotbaby 使用了53k的高质量动漫图片,在A40 48G 上训练超过了 150个 GPU小时。

在训练集方面,使用了相当部分的裸体训练集,作者相信,这些内容本应是一个健康的大模型应该具备的,作者希望通过这种学习,模型可以更好地理解人体结构。

当不指定服装或者直接使用角色名称时,会有大概率生成裸露的图片,因此,强烈建议您在公共场所谨慎使用此模型。

训练配置

在提示词的具体使用上,可以参考Animagine XL 3.0使用的负面提示,以引导模型生成高美观的图像

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name

为了获得更高质量的结果,请在提示前添加以下内容:

masterpiece, best quality

训练配置

在训练参数上,此模型大部分参考了Animagine XL 3.0的训练配置

但在学习率和是否训练文本编码器上,做了少许调整。

鸣谢名单

特别鸣谢:算力赞助 / GPU Sponsor: nieta.art

最尊敬的模型团队:cagliostrolab

训练集提供:二小姐

感谢在模型过程中提供指导和帮助的前辈:

@kblueleaf

@L_A_X

@Euge

@LEOSAM

鸣谢社群:二次元的XL开放乐园

鸣谢团队:NAIXL

特别提醒

未经作者许可,请勿将此模型转载或以任何方式发布到其他平台!

由此产生的一切不良后果,由转载者以及对应平台承担。

附:训练参数(Training parameters)

[sdxl_arguments]
cache_text_encoder_outputs = false
no_half_vae = false
min_timestep = 0
max_timestep = 1000

[model_arguments]
pretrained_model_name_or_path = "/root/autodl-tmp/stable-diffusion-webui/models/Stable-diffusion/animagine-xl-3.0.safetensors"

[dataset_arguments]
shuffle_caption = true
debug_dataset = false
train_data_dir = "/root/autodl-tmp/53k"
dataset_repeats = 1
keep_tokens_separator = "|||"
resolution = "1024, 1024"
caption_dropout_rate = 0
caption_tag_dropout_rate = 0
caption_dropout_every_n_epochs = 0
token_warmup_min = 1
token_warmup_step = 0
enable_bucket = true
min_bucket_reso=640 
max_bucket_reso=2048
bucket_reso_steps=64

[training_arguments]
output_dir = "/root/autodl-tmp/stable-diffusion-webui/models/Stable-diffusion/53k"
output_name = "chenkin_53kv2"
save_precision = "fp16"
train_batch_size=6         
vae_batch_size=4 
max_train_epochs=10 
save_every_n_epochs=1
max_token_length = 225
mem_eff_attn = false
xformers = true
sdpa = false
max_data_loader_n_workers = 8
persistent_data_loader_workers = true
gradient_checkpointing = true
gradient_accumulation_steps = 1
mixed_precision = "fp16"

[sample_prompt_arguments]
sample_every_n_steps = 1000
sample_sampler = "euler_a"
sample_prompts="/root/example.txt"

[saving_arguments]
save_model_as = "safetensors"

[optimizer_arguments]
optimizer_type = "AdaFactor"
learning_rate = 7.5e-7
train_text_encoder = false
learning_rate_te1 = 0
learning_rate_te2 = 0
optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False",]
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 100
max_grad_norm = 0