v9

注意：这个版本尝试额外添加了一个触发词“ppw”。把它放在提示词的开始处来触发lora。
Note: This version experimentally adds an extra trigger word: “ppw”. Place it at the beginning of your prompt to trigger the LoRA.

v10版本对过去的数据集进行了维护，利用一系列AI工具修复了100+张过去曾因为“工作量太大难以修复”而删除的图片，并且添加了一些实新合成的数据。现在使用触发词“ppw”应该会得到比之前更强烈的风格。
The v10 version includes a maintenance of the previous dataset. With the help of various AI tools, restored 100+ images that had previously been removed because of “requiring too much work to fix.” Added some newly synthesized data. Using the trigger word “ppw” should now produce a stronger stylization than before.
在v9版本中，我尝试添加了一个约2000张的真人照片数据集，让模型拥有生成不同族裔的画像的能力。但是经过大量测试后，我发现真人照片会显著地降低二次元模型的图片清晰度。这个版本降低了写实照片的训练轮数，尽管对于族裔标签的响应效果弱于v9，但是提升了生成质量。族裔相关的tag将在未来尝试使用别的方式来加强。
In v9, I tried adding a real-photo dataset of about 2,000 images to give the model the ability to generate portraits of different ethnicities. However, after extensive testing, I found that real photos significantly reduce image clarity in anime-style models. In this version, I reduced the number of training epochs for the photo dataset. Although the response to ethnicity tags is weaker than in v9, the overall generation quality has improved. Ethnicity-related tags will be strengthened through other methods in the future.
新增一个族裔标签Anime。使用它可以让人物获得更接近传统二次元的脸型。
Added a new ethnicity tag: Anime. Using this tag can give characters a face shape closer to traditional anime.
经过近一年的尝试，我认为通过训练很难消除SDXL的某些固有缺陷。这很可能是由模型结构、vae等共同决定的。使用这些素材反而增加了许多生成结果的不确定性。移除了质量提示词“low quality”和“best quality”的训练数据集。
After nearly a year of experimentation, I believe it is difficult to eliminate some of SDXL’s inherent flaws through training alone. These issues are most likely determined by a combination of factors such as the model architecture and VAE. Including this type of material instead introduced more uncertainty into the generated results. Therefore, the training datasets for the quality prompt tags “low quality” and “best quality” have been removed.
降低了学习率，增加训练步数，缓解了一些人物解剖错误、肢体扭曲的问题。
The learning rate was reduced and the number of training steps was increased, which helps alleviate some anatomy errors and limb distortion issues.

快速上手 | Quick Start

这是什么？ | What is this?

People's Works是一个实验性的微调模型系列，最初基于一个由Pony V6 XL生成的图片构成的数据集。这个模型的数据集由数千张AI社区用户发布的图片和作者使用AI合成的图片构成，经过人工筛选、编辑、修改和标注后用于训练。除此之外还有一个超过2000张由照片、游戏CG和3d渲染图像构成的辅助数据集，用于额外补充知识。
People’s Works is an experimental fine-tuned model series, originally based on a dataset composed of images generated by Pony V6 XL. The dataset consists of several thousand images published by AI community users, along with images synthesized by the author using AI. These images were manually curated, edited, modified, and annotated before being used for training. In addition, there is an auxiliary dataset of over 2,000 images composed of photographs, game CGs, and 3D renders, used to provide supplementary knowledge.

模型功能 | Model features

本系列模型主要功能是帮助基础模型在不使用画师串、较少的质量提示词的条件下获得相对稳定的风格化图像，为提示词节约token空间。
使用人工挑选和标注的数据集强化训练正面和负面审美提示词。
模型针对flat color、realistic等特定风格肌理进行了强化。
对人物的年龄、族裔、肤质等特征的精细化控制。
使用了较高的训练分辨率，使基础模型在高清修复时表现更好。
通过手工修复图片的方式降低了模型的某些细节瑕疵出现的概率。
The primary function of this series of model is to help the basemodel generate relatively stable, stylized images, without artist keywords or long quality tags, freeing up token space for prompts.
By using manually selected and annotated datasets, the model strengthens both positive and negative aesthetic tags training.
The model includes targeted enhancements for specific visual textures and styles, such as flat color and realistic.
It enables finer control over character attributes, including age, ethnicity, and skin texture.
A higher training resolution is used, improving the basemodel’s performance during high-res upscaling.
By manually editing the images, the likelihood of certain flaws appearing in the model’s outputs has been reduced.

使用方法 | Usage

v10

positive:

ppw, masterpiece, very aesthetic

negative: /

v7-v9

positive:

masterpiece, best quality, very aesthetic

negative:

low quality, displeasing

更新记录 | Change log

v9

更改了系列名称。自这个版本起，训练集中来自Pony v6 XL的图片已经在所有AIGC内容中占比不足1/3。随着越来越多的新模型出现，我有计划在明年将这个系列拓展到其他模型上。为了避免未来用户使用上的混乱和误解，这个系列从这个版本起更改命名。
The series name has been changed. Starting from this version, images sourced from Pony v6 XL make up less than one third of the training data across all AIGC content. As increasing number of new models are emerging, I plan to expand this series to other models next year. To avoid potential confusion and misunderstanding for users in the future, the series name has been changed starting from this version.
这个版本的训练方式是直接训练LoCon，而非训练Checkpoint后再抽取LoRA。模型相较前一个版本效果更强。
This version is trained directly as a LoCon, rather than training a checkpoint first and then extracting a LoRA. Compared to the previous versions, the model delivers stronger effects.
v9全系列使用1536分辨率的训练集。现在使用这个Lora生成图片时支持单边768-1536的分辨率，使用高清修复时也可以尝试更高的denoise参数了。
All v9 models use a 1536-resolution training set. When generating images with this LoRA, single-side resolutions from 768 to 1536 are now supported. When using high-res fix, you can also try higher denoise values.
对训练图片调色。现在模型在没有指定色彩时更倾向于生成暖色调的图片，并且色彩的饱和度略微提高。较暗的场景明暗对比更大了。
Color adjustments were applied to images. When no specific color is specified, the model now tends to produce warmer tones, with slightly increased saturation. Darker scenes also have stronger contrast between light and shadow.
之前版本的数据集中，人物鼻子的画法不统一。出于作者本人的兴趣，手工修改了其中约300幅图片，并暂时排除了约200幅来不及修改的图片。现在人物的鼻子有鼻翼了。
In earlier versions of dataset, nose depiction was inconsistent. Out of personal interest, the author manually modified around 300 images and temporarily excluded about 200 images that could not be edited in time. Characters’ noses now have nose wings.
删除了旧版本中数百张过时的低质量训练数据。
Hundreds of outdated, low-quality images from older versions of dataset have been removed.
添加了一个新的实验性数据集：使用真人相片作为引导，现在你可以使用以下年龄和族裔标签了：
A new experimental dataset has been added. Using real photographs as guidance, you can now use the following age and ethnicity tags:

child, teenage, adult, mature

Caucasian, Asian, Indian, African

我的标签设计优先选择Danbooru数据集中已经存在的标签，尽管其中很多只有很少量的数据，在原版模型中几乎无法触发。重新启用了已经被删除的danbooru词条Caucasian和teenage，增设了adult和African两个标签。此外，loli和shota因为其文化背景中强烈的性暗示倾向，这两个词条被完全替换，根据具体情况分流入child和teenage。
My tag design prioritizes labels that already exist in the Danbooru dataset, even though many of them had very limited data and are therefore almost impossible to trigger in the base models. The previously removed Danbooru tags Caucasian and teenage have been re-enabled, and two new tags, adult and African, have been added. Additionally, due to the strong sexual connotations of loli and shota in their cultural context, these tags have been completely replaced and redistributed into child and teenage depending on the situation.

v8

肌理更新：强化了以下tag的学习：

Texture Update: The following tags have been reinforced in training:

realistic, photorealistic, flat color,

shiny skin, matte skin, shiny hair,

请注意，在danbooru数据集中有很多个用于描述“照片”和“接近照片的风格”的tag。我在训练集中统一标注这些图片为“photorealistic”。但是使用danbooru训练集训练的SDXL模型大多并不能很好地画出写实图像，因此“photorealistic”只建议在较小权重下用于改变画面的肌理。“realistic”可以在高权重下正常运作。

Please note that Danbooru dataset contains multiple tags to describe "photo" or "photo-like styles". I’ve tagged all such images as “photorealistic” in dataset.

However, most SDXL models trained on the Danbooru dataset do not render realistic images well. “photorealistic” is only recommended at low weight, where it can help adjust texture rather than create realism images. The “realistic” tag can work properly at higher weight.

v8版本简介：

Pony: People's Works （ppw）是一个实验性的微调模型系列，数据集有约85%是收集自CivitAI上用户发表的AI生成图片。早期ppw的数据集最初建立在由pony v6生成图片的基础上，因此本系列模型生成的图片也带有pony diffusion的特征。
本系列模型使用标准Danbooru标签，主要擅长生成中、近景风格化人像。它们的主要功能是使基础模型可以在不使用画师串、较少的质量提示词的条件下获得相对稳定的图像质量，为提示词节约token空间。
本模型并非风格模型，在不同的提示词和生成条件下可能会有微妙的画风差异。

Pony: People's Works (ppw) is a experimental fine-tuned model series, approximately 85% of the dataset comes from AI-generated images published by users on CivitAI. Since the earlier ppw dataset was built on images generated by Pony V6, the outputs of this series also carry some characteristics of Pony Diffusion.
This series uses standard Danbooru tags and is mainly optimized for generating stylized portraits at medium and close range. The primary effect of this model series is to allow the basemodel to achieve relatively stable image quality, without artist keywords or long quality tags, freeing up token space for prompts.
These models are not style LoRAs. There may be subtle stylistic variations depending on different prompts and generating conditions.

v7

v7版本对数据集结构做了较大幅度的调整，并且使用了不同的训练参数和训练策略，因此可能v7会不如原来的版本稳定。

The v7 version has undergone significant structural adjustments to the dataset, and utilizes different training parameters and strategies. As a result, v7 may be less stable than the previous versions.

v-pred模型在civitAI在线生成器上的表现和吐司的在线生成表现完全不一样，相同参数完全无法复现。我也不知道是为什么......

The v-pred model's performance on the CivitAI online generator is completely different from online generation on TensorArt. The results are entirely unreproducible with a same parameters. I have no idea why...

TensorArt version CivitAI ver. with same parameter on CivitAI with higher weight

v7版本简介：

这是一个在前作的数据集基础上发展而来的图像质量LoCon，约90%-95%图片数据来自CivitAI上发布的图片。

它使模型可以在不使用画师串、使用较少的质量提示词的条件下获得相对稳定的图像质量，节约出更多的token空间，同时它还可以修复一部分模型固有的生成瑕疵。（但是不包括手部）

因为数据集选取的原因，生成的图片会带有Pony的质感。但因为它并不指向任何特定的画师、风格和绘画技法，所以在不同的提示词、模型条件下可能会有微妙的画风差异。

This is a generation quality LoCon developed based on the dataset from the previous work. About 90%-95% of the image data comes from CivitAI.

It allows models to achieve relatively stable image quality without artist tags or using long quality prompts, freeing up more token space. Additionally, it can fix some inherent generation flaws of the model. (except for hands)

Due to the dataset selection, the generated images exhibit a Pony-like style. However, since it does not reference any specific artist, style, or painting technique, there may be subtle stylistic variations depending on different prompts and checkpoint conditions.

数据集来源及许可证 | Dataset Source & License

数据集中每一张图片都经过作者本人的人工筛选、分类和标注编辑，其中上千张图片经过人工的编辑、对细节瑕疵进行修正。
此模型为免费、开源模型，用户可以在私人设备上自行部署该模型。作者并不从模型出售中获取任何报酬。作者并不限制本系列模型用于商业生成服务或者生成图像用于商业用途，但是请注意配合使用的Checkpoint和其他LoRA的许可证限制。
请注意本模型的数据集由一个为约5000张AI图像构成的训练数据集和一个超过2000张图像构成的辅助数据集组成。其中，主数据集的图片大部分收集自AI社区。辅助数据集则包括公开的新闻图片、游戏CG和宣传图、3D渲染图和已购买的商业写真等。辅助数据集仅用于学习光影色彩、构图术语、人体解剖特征等一般通用知识，使用本模型不能还原辅助数据集中的版权内容。当前法律对这类数据的使用没有明确的统一规定，请有商用意向的本系列模型用户自行注意相关风险。
本数据集没有训练任何独立画师的数据，也没有标注任何画师ID信息（不排除AI错误标注的情况）。
另外，本模型不允许用作闭源商用、模型出售，也禁止用于闭源商用模型的融合。对于开源融合模型用于生成服务的情形不做限制，但是建议标注融合模型的出处。

Every image in the dataset was manually reviewed, categorized, and annotated by the author. Among them, more than a thousand images were additionally hand-edited to manually correct fine-detail visual defects.
This model is free and open-source model, allowing users to deploy it on their personal devices. The author does not receive any compensation from selling the model. The author does not impose restrictions on using this model for commercial image generation services or generating images for commercial purposes. However, please be mindful of the license restrictions of the Checkpoint and other LoRAs used alongside this model.
Please note that this model’s dataset consists of a primary training set of approximately 5,000 AI-generated images and an auxiliary dataset of over 2,000 images. The majority of images in the main dataset were collected from AI communities. The auxiliary dataset includes publicly available news photographs, game CGs and promotional images, 3D-rendered images, and purchased photo sets.
The auxiliary dataset is used solely to learn general-purpose knowledge such as lighting, composition terminology, and human anatomical features. This model cannot reproduce or restore copyrighted content from the auxiliary dataset. As current laws do not provide a clear, unified standard for the use of such data, users who intend to use this model for commercial purposes should be aware of and assess the associated legal risks on their own.
This dataset does not include training data from any individual artist, nor does it contain explicit artist attributions (though AI mistagging cannot be entirely ruled out).
Additionally, this model is not permitted for use in closed-source commercial applications, model resales, or merged into closed-source commercial models. There are no restrictions on open-source merged models being used for image generation services, but it is recommended to credit the sources of any merged models.

Files with SHA256: d55377bfb9ba284133f3a7cfa003f29845a605174f54df8a6d314e507bc41f9d