<h1 id="creation-naixl-v1-251001">Creation NAIXL V1-251001</h1><hr />中文 / English<h2>简介</h2>这一模型基于 NoobAI-XL 的 <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1116447">Epsilon 系列模型</a>，使用全新的数据集进行训练，希望你能喜欢。我会在此基础上扩大数据集规模继续训练，直至其完善。十月版本的模型对比五月在多方面得到了显著增强，但语义准确性降低了不少，有待继续改进。<h2>模型训练详情</h2><h3>数据集</h3>继续训练所使用的数据集的生成样本主要是从 Pixiv 收集的图像，提示词则包括某一图像在 Pixiv 由作者或群众添加的英文标签，以及由 <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/wiki_pages/help:api">Danbooru API</a> 获取的标签作为提示词。对于未上传到 Danbooru 的图像，由于 Pixiv 对作品提供的标签较少，故利用 <code>wd-tagger-v3</code> 对每一个图像推理，进一步获取作为提示词的 danbooru 标签，然后使用本地的索引，配合 Danbooru 提供的自动补全 API，映射每一个画师的用户名为正确的 danbooru 艺术家标签。详情如下：从 Pixiv 的<a target="_blank" rel="ugc" href="https://www.pixiv.net/ranking.php?mode=monthly">常规月排行榜</a>完整爬取从 2025 年 1 月至 10 月每天的 500 张图像后去重，并获取每张图像的英文标签。从 Pixiv 的<a target="_blank" rel="ugc" href="https://www.pixiv.net/ranking.php?mode=weekly_r18">R18周排行榜</a>完整爬取从 2025 年 1 月至 5 月每天的图像后去重，并获取每张图像的英文标签，之所以没有继续更新 R18 排行榜上的图片是因为打码的太多了。从 Pixiv VISIONS 年鉴中选取约 50 名插画作者从 2024 年 11 月 1 日至 2025 年 10 月 1 日的作品，并获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 作品被审查过，所以从 <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/">Danbooru</a> 按 <code>order:score age:&lt;1year</code> 遴选近 7000 张正面样本，涵盖多数游戏角色，并收集负面样本 <code>bad hands, bad feet...</code>。合计 21,345 张样本。<pre><code>WEEKLY_RANKING_R18_API: str = "https://www.pixiv.net/ranking.php?mode=weekly_r18&amp;date={date}&amp;p={p}&amp;format=json"

MONTHLY_RANKING_API: str = "https://www.pixiv.net/ranking.php?mode=monthly&amp;date={date}&amp;p={p}&amp;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}&amp;page={page}&amp;limit={limit}"

DANBOORU_AUTO_COMPLETE_API: str = "https://danbooru.donmai.us/autocomplete.json"</code></pre>对提示词去重，删去部分错误提示词，删去提示词过短或过长的训练样本，删去含以下提示词的样本。<pre><code>chat log, fake screenshot, ai-generated, ai-generated illustration, announcement celebration, comic, manga, how to draw, multiple boys, multiple girls</code></pre>按作品的热度自动添加质量标签（见下文）。<h2>生成指引</h2><h3>提示词</h3>在每次生成中，你应该统一使用同一种写法的提示词，可供使用的提示词书写方式有两种。<ul><li>移除提示词中的所有下划线，如 <code>robin \(honkai: star rail\)</code> （建议，因为训练用的提示词格式就是这样的）。</li></ul><ul><li>含下划线的提示词，如 <code>robin_\(honkai:_star_rail\)</code>；</li></ul>推荐的提示词顺序如下：<pre><code>&lt;1boy/1girl/1other/...&gt;, &lt;character&gt;, &lt;artists/styles&gt;, &lt;quality tags&gt;, &lt;composition tags&gt;, &lt;IP/franchise&gt;, &lt;more tags&gt;</code></pre>数据集中将 Pixiv 中的 “xxx n+ bookmarks/users” 标签映射为了质量提示词，以此增强了模型的质量控制。经对比，这些提示词的效果已较为显著。可用的质量提示词有：<pre><code>masterpiece # n &gt;= 10000
best quality # 5000 &lt;= n &lt; 10000
high quality # 1000 &lt;= n &lt; 5000
good quality # 500 &lt;= n &lt; 1000
normal quality # else</code></pre>延续了 NAI-XL 的年代提示词，现在你可以使用 <code>year 2025</code>，且 <code>newest</code> 对应的年份更改为 2021~2025 。我还对图像的分辨率作了如下划分：<pre><code>absurdres # n &gt; 9,000,000 (pixels)
highres # 4,000,000 &lt; n &lt;= 9,000,000 (pixels)
midres # 1,048,576 &lt; n &lt;= 4,000,000 (pixels)
lowres # else</code></pre>由于数据集的构成，你可以使用不限于 Danbooru 的、存在于 Pixiv 标签中的更多提示词。尽管它们目前的效果还不明显，但我相信，模型经过后续的迭代，会使它们趋于完善。我在用的负面提示词如下：<pre><code>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</code></pre><h3>其他</h3>采样器：Euler / Euler aCFG：4.0~6.5迭代步数：主要阶段应大于 20 步未提到的部分与原模型保持一致。<hr /><h2 id="introduction">Introduction</h2>This model is based on the <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1116447">Epsilon series</a> 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.<h2 id="model-training-details">Model Training Details</h2><h3 id="dataset">Dataset</h3>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 <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/wiki_pages/help:api">Danbooru API</a>.For images not uploaded to Danbooru — since Pixiv generally provides fewer tags — I used <code>wd-tagger-v3</code> 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:<ul><li>Crawled the <a target="_blank" rel="ugc" href="https://www.pixiv.net/ranking.php?mode=monthly">Pixiv monthly ranking (regular)</a> from January to October 2025, collecting 500 images per day, deduplicated, and extracted English tags for each image.</li><li>Crawled the <a target="_blank" rel="ugc" href="https://www.pixiv.net/ranking.php?mode=weekly_r18">Pixiv R18 weekly ranking</a> from January to May 2025, deduplicated and extracted English tags for each image. (Did not continue updating R18 rankings due to excessive censorship.)</li><li>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.</li><li>Due to censorship policies in Pixiv’s host country, most R18 works are moderated, so an additional 7,000 positive samples were selected from <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/">Danbooru</a> under <code>order:score age:&lt;1year</code>, covering most popular game characters. Negative samples were also collected (e.g., bad hands, bad feet, etc.).</li></ul>Total: 21,345 samples.<pre><code>WEEKLY_RANKING_R18_API: str = "https://www.pixiv.net/ranking.php?mode=weekly_r18&amp;date={date}&amp;p={p}&amp;format=json"

MONTHLY_RANKING_API: str = "https://www.pixiv.net/ranking.php?mode=monthly&amp;date={date}&amp;p={p}&amp;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}&amp;page={page}&amp;limit={limit}"

DANBOORU_AUTO_COMPLETE_API: str = "https://danbooru.donmai.us/autocomplete.json"</code></pre>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:<pre><code>chat log, fake screenshot, ai-generated, ai-generated illustration, announcement celebration, comic, manga, how to draw, multiple boys, multiple girls</code></pre>Quality tags were automatically assigned according to the artwork’s popularity (see below).<h2 id="generation-guidelines">Generation Guidelines</h2><h3 id="prompts">Prompts</h3>During generation, you should keep a consistent writing style for prompts. Two formats are supported:<ul><li>Without underscores: <code>robin \(honkai: star rail\)</code></li><li>With underscores: <code>robin_\(honkai:_star_rail\)</code></li></ul>Recommended prompt order:<pre><code>&lt;1boy/1girl/1other/...&gt;, &lt;character&gt;, &lt;artists/styles&gt;, &lt;quality tags&gt;, &lt;composition tags&gt;, &lt;IP/franchise&gt;, &lt;more tags&gt;</code></pre>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:<pre><code>masterpiece # n &gt;= 10000
best quality # 5000 &lt;= n &lt; 10000
high quality # 1000 &lt;= n &lt; 5000
good quality # 500 &lt;= n &lt; 1000
normal quality # else</code></pre>The year tag system from NAI-XL is continued — you can now use <code>year 2025</code>, and <code>newest</code> now corresponds to the range 2021–2025.I also categorized images by resolution as follows:<pre><code>absurdres # n &gt; 9,000,000 (pixels)
highres # 4,000,000 &lt; n &lt;= 9,000,000 (pixels)
midres # 1,048,576 &lt; n &lt;= 4,000,000 (pixels)
lowres # else</code></pre>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.<pre><code>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</code></pre><h3 id="other-settings">Other Settings</h3><ul><li>Sampler: Euler / Euler a</li><li>CFG: 4.0–6.5</li><li>Steps: Should exceed 20 during main generation phases</li><li>Unmentioned parts remain consistent with the base model.</li></ul>Translated by GPT-5.

Files with SHA256: 503a09a4895dabb76992b5769425c35adc7cc788d756db62e6a2187a253c128c