Creation NAIXL / 2025-Oct

ChanterChanter
99
794
动画base modelstyle artcheckpoint trained
eps-v1-251001eps-v1-250513eps-v1-preview

详情

下载文件

模型描述

Creation NAIXL V1-251001

中文 / English

简介

本模型基于 NoobAI-XL 的 Epsilon 系列模型，使用全新数据集进行训练，希望您会喜欢。我将持续扩大数据集并继续训练，直至模型趋于完善。与五月版本相比，十月版本在多个方面均有显著提升，但语义准确性有所下降，仍需进一步改进。

模型训练详情

数据集

用于继续训练的数据样本主要来自 Pixiv，提示词包括图像在 Pixiv 上由作者或用户添加的英文标签，以及通过 Danbooru API 获取的标签。

对于未上传至 Danbooru 的图像，由于 Pixiv 提供的标签较少，因此使用 wd-tagger-v3 对每张图像进行推理，进一步获取作为提示词的 Danbooru 标签；随后通过本地索引，结合 Danbooru 提供的自动补全 API，将每位画师的用户名映射为正确的 Danbooru 艺术家标签。

具体如下：

从 Pixiv 的常规月排行榜中完整爬取 2025 年 1 月至 10 月每日前 500 张图像，去重并提取每张图像的英文标签。

从 Pixiv 的R18周排行榜中完整爬取 2025 年 1 月至 5 月每日图像，去重并提取每张图像的英文标签。未继续更新 R18 排行榜图像的原因是大量内容被打码。

从 Pixiv VISIONS 年鉴中选取约 50 名插画师于 2024 年 11 月 1 日至 2025 年 10 月 1 日期间的作品，并提取每张图像的英文标签。

由于 Pixiv 所在国家的政策原因，绝大多数 R18 作品均经过审查，因此从 Danbooru 中按 order:score age:<1year 筛选近 7000 张正面样本，涵盖多数游戏角色，并收集负面样本如 bad hands, bad feet...。

总计 21,345 张样本。

WEEKLY_RANKING_R18_API: str = "https://www.pixiv.net/ranking.php?mode=weekly_r18&date={date}&p={p}&format=json"

MONTHLY_RANKING_API: str = "https://www.pixiv.net/ranking.php?mode=monthly&date={date}&p={p}&format=json"

ILLUST_INFO_API: str = "https://www.pixiv.net/ajax/illust/{illust_id}?lang=en"

DANBOORU_SEARCH_API: str = "https://danbooru.donmai.us/posts.json?tags={tags}&page={page}&limit={limit}"

DANBOORU_AUTO_COMPLETE_API: str = "https://danbooru.donmai.us/autocomplete.json"

对提示词进行去重，删除错误提示词，剔除提示词过短或过长的样本，以及包含以下提示词的样本：

chat log, fake screenshot, ai-generated, ai-generated illustration, announcement celebration, comic, manga, how to draw, multiple boys, multiple girls

根据作品热度自动添加质量标签（见下文）。

生成指引

提示词

每次生成时，应统一使用同一种提示词写法，目前支持两种格式：

  • 去除所有下划线，如 robin \(honkai: star rail\)（推荐，因为训练数据提示词即为此格式）。

  • 保留下划线，如 robin_\(honkai:_star_rail\)；

推荐的提示词顺序如下：

<1boy/1girl/1other/...>, <character>, <artists/styles>, <quality tags>, <composition tags>, <IP/franchise>, <more tags>

数据集中将 Pixiv 中的 “xxx n+ 收藏/用户数” 标签映射为质量提示词，以增强模型的质量控制。经测试，这些提示词效果已相当显著。

可用的质量提示词包括：

masterpiece         # n >= 10000
best quality        # 5000 <= n < 10000
high quality        # 1000 <= n < 5000
good quality        # 500 <= n < 1000
normal quality      # else

延续 NAI-XL 的年代提示词系统，现在可使用 year 2025，且 newest 对应的年份范围已更新为 2021~2025。

我还对图像分辨率进行了如下划分：

absurdres           # n > 9,000,000 像素
highres             # 4,000,000 < n <= 9,000,000 像素
midres              # 1,048,576 < n <= 4,000,000 像素
lowres              # else

由于数据集构成，您可使用 Pixiv 标签中存在但不属于 Danbooru 的更多提示词。尽管目前效果尚不明显，但我相信经过后续迭代，这些提示词将逐步完善。

我常用的负面提示词如下：

Suggest Negative Prompt:
lowres, (worst quality, bad quality, low quality:1.2), bad anatomy, bad perspective, bad hands, bad feet, bad pixiv id, anime screencap, watermark, artist name, censored, bar censor, mosaic censoring, amputee

其他

采样器：Euler / Euler a

CFG：4.0~6.5

迭代步数：主要生成阶段应超过 20 步

未提及部分与原模型保持一致。

Introduction

This model is based on the Epsilon series from NoobAI-XL and was trained using a brand-new dataset. I hope you’ll like it. I will continue expanding the dataset and training the model until it becomes fully refined. Compared to the May version, the October version has seen significant improvements in multiple aspects. However, the semantic accuracy has decreased significantly and needs to be further improved.

Model Training Details

Dataset

The samples used for continued training were primarily collected from Pixiv.
Each image’s prompts include English tags added by the author or the community on Pixiv, along with additional tags retrieved via the Danbooru API.

For images not uploaded to Danbooru — since Pixiv generally provides fewer tags — I used wd-tagger-v3 to infer Danbooru-style tags for each image.
Then, using a local index and Danbooru’s autocomplete API, each artist’s Pixiv username was mapped to the correct Danbooru artist tag.

Details are as follows:

  • Crawled the Pixiv monthly ranking (regular) from January to October 2025, collecting 500 images per day, deduplicated, and extracted English tags for each image.

  • Crawled the Pixiv R18 weekly ranking from January to May 2025, deduplicated and extracted English tags for each image. (Did not continue updating R18 rankings due to excessive censorship.)

  • Selected works from about 50 illustrators featured in Pixiv VISIONS Yearbook, covering their works from Nov 1, 2024 to Oct 1, 2025, with English tags retrieved for each image.

  • Due to censorship policies in Pixiv’s host country, most R18 works are moderated, so an additional 7,000 positive samples were selected from Danbooru under order:score age:<1year, covering most popular game characters.
    Negative samples were also collected (e.g., bad hands, bad feet, etc.).

Total: 21,345 samples.

WEEKLY_RANKING_R18_API: str = "https://www.pixiv.net/ranking.php?mode=weekly_r18&date={date}&p={p}&format=json"

MONTHLY_RANKING_API: str = "https://www.pixiv.net/ranking.php?mode=monthly&date={date}&p={p}&format=json"

ILLUST_INFO_API: str = "https://www.pixiv.net/ajax/illust/{illust_id}?lang=en"

DANBOORU_SEARCH_API: str = "https://danbooru.donmai.us/posts.json?tags={tags}&page={page}&limit={limit}"

DANBOORU_AUTO_COMPLETE_API: str = "https://danbooru.donmai.us/autocomplete.json"

Duplicate prompts were removed, incorrect ones deleted, and samples with overly short or long prompts were discarded.
Samples containing any of the following prompt tags were also removed:

chat log, fake screenshot, ai-generated, ai-generated illustration, announcement celebration, comic, manga, how to draw, multiple boys, multiple girls

Quality tags were automatically assigned according to the artwork’s popularity (see below).

Generation Guidelines

Prompts

During generation, you should keep a consistent writing style for prompts. Two formats are supported:

  • Without underscores: robin \(honkai: star rail\)

  • With underscores: robin_\(honkai:_star_rail\)

Recommended prompt order:

<1boy/1girl/1other/...>, <character>, <artists/styles>, <quality tags>, <composition tags>, <IP/franchise>, <more tags>

Pixiv’s “xxx n+ bookmarks/users” tags were mapped to quality prompt tags to improve quality control.
After testing, these tags have shown significant effects.

Available quality tags:

masterpiece         # n >= 10000
best quality        # 5000 <= n < 10000
high quality        # 1000 <= n < 5000
good quality        # 500 <= n < 1000
normal quality      # else

The year tag system from NAI-XL is continued — you can now use year 2025,
and newest now corresponds to the range 2021–2025.

I also categorized images by resolution as follows:

absurdres           # n > 9,000,000 (pixels)
highres             # 4,000,000 < n <= 9,000,000 (pixels)
midres              # 1,048,576 < n <= 4,000,000 (pixels)
lowres              # else

Because of how the dataset is constructed, you can use not only Danbooru tags but also many Pixiv-specific tags.
Their current impact is limited, but with future iterations, they should become more effective.

Suggest Negative Prompt:
lowres, (worst quality, bad quality, low quality:1.2), bad anatomy, bad perspective, bad hands, bad feet, bad pixiv id, anime screencap, watermark, artist name, censored, bar censor, mosaic censoring, amputee

Other Settings

  • Sampler: Euler / Euler a

  • CFG: 4.0–6.5

  • Steps: Should exceed 20 during main generation phases

  • Unmentioned parts remain consistent with the base model.

Translated by GPT-5.

此模型生成的图像

未找到图像。