Creation NAIXL

Creation NAIXL V1中文 / English简介这一模型基于 NoobAI-XL 的 Epsilon 系列模型，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。模型训练详情数据集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标签，以及由模型自动推理得到的 danbooru 作为提示词。详情如下：从 Pixiv 的常规月排行榜完整爬取从 2025 年 1 月至 5 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R18周排行榜完整爬取从 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。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 5 月 10 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 Patreon 与 Pixiv Fanbox 收集了部分未审查的作品。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"利用本地的索引，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。由于 Pixiv 对作品提供的标签较少，故利用 DeepDanbooru 模型，以75%置信度对每一个图像推理，进一步获取作为提示词的 danbooru 标签。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。chat log, fake screenshot, ai-generated, ai-generated illustration, announcement celebration, comic, manga, how to draw, multiple boys, multiple girls按作品的热度自动添加质量标签（见下文）。训练使用 AdamW 8bit 优化器，使用余弦退火学习率衰减，在 RTX 4090 上进行训练，共迭代约 274,000 步。模型损失率整体呈下降趋势，过程中未遇到需要手动回退检查点的情况。生成指引提示词在每次生成中，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。含下划线的提示词，如 robin_(honkai:_star_rail)；移除提示词中的所有下划线，如 robin (honkai: star rail) 。推荐的提示词顺序如下：<artists/styles>, <quality tags>, <composition tags>, <IP/franchise>, <character>, <1boy/1girl/1other/...>, <more tags>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：masterpiece # n >= 10000 best quality # 5000 <= n < 10000 high quality # 1000 <= n < 5000 good quality # 500 <= n < 1000 normal quality # else延续了 NAI-XL 的年代提示词，现在你可以使用 year 2025，且 newest 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：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：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 aCFG：4.0~6.5迭代步数：主要阶段应大于28步未提到的部分与原模型保持一致。IntroductionThis model is based on the Epsilon series of NoobAI-XL and has been trained on an entirely new dataset, I hope you find it enjoyable and effective. I will continue to expand the dataset and refine the model until it reaches a more complete and mature state.Model Training DetailsDatasetThe samples used for further training were generated primarily from images collected on Pixiv. The prompts include both the tags added by the original artist or community on Pixiv and automatically inferred Danbooru tags. Details are as follows:A complete crawl of the top 500 images from Pixiv's monthly rankings, covering every day from January to May 2025, was performed. After removing duplicates, English tags for each image were extracted.A complete crawl of images from Pixiv’s R18 weekly rankings for each day from January to May 2025 was also performed, followed by deduplication and extraction of English tags.From the Pixiv VISIONS annual anthology, selected works by about 50 illustrators between November 1, 2024 and May 10, 2025, retrieving each image’s English tags.Because of local censorship in Pixiv’s host country, most R18 works are censored; therefore, additional uncensored works were gathered from Patreon and Pixiv Fanbox.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"Mapped each artist’s Pixiv username to the correct Danbooru artist tag using a local index and Danbooru’s autocomplete API.Since Pixiv tags are relatively sparse, we employed the DeepDanbooru model (with a 75% confidence threshold) to infer additional Danbooru tags for each image.Deduplicated prompts, removed obviously incorrect tags, discarded training samples with prompts that were too short or too long, and filtered out any samples containing undesirable tags such as:chat log, fake screenshot, ai-generated, ai-generated illustration, announcement celebration, comic, manga, how to draw, multiple boys, multiple girlsAutomatically assigned quality tags based on each work’s popularity (see below).TrainingThe model was trained using the AdamW 8-bit optimizer with cosine annealing learning rate decay. Training was conducted on an RTX 4090 GPU for approximately 274,000 iterations. The overall loss showed a consistent downward trend, and no manual checkpoint rollbacks were required during the training process.Generation GuidelinesPromptsWhen generating images, use a consistent prompt format. Two acceptable styles are:1. Underscore style, e.g.robin_(honkai:_star_rail)2. Space style (no underscores), e.g.robin (honkai: star rail)Recommended tag order:<artists/styles>, <quality tags>, <composition tags>, <IP/franchise>, <character>, <1boy/1girl/1other/...>, <more tags>The dataset maps Pixiv’s “xxx n+ bookmarks/users” tags to quality prompts to enhance quality control. Comparative testing shows these tags are effective. Available quality tags:masterpiece # n >= 10000 best quality # 5000 <= n < 10000 high quality # 1000 <= n < 5000 good quality # 500 <= n < 1000 normal quality # elseThe temporal prompt conventions from NAI-XL have been retained. You may now use the term year 2025, and the range represented by newest has been updated to encompass the years 2021 through 2025.Additionally, I have introduced the following classifications based on image resolution: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 # elseBecause our dataset includes a variety of Pixiv tags beyond Danbooru, you may experiment with additional Pixiv tags. Their effects are still emerging, but should improve as the model evolves.The negative prompt words I'm using are as follows: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, amputeeOther SettingsSampler: Euler / Euler aCFG scale: 4.0~6.5Steps: typically > 28 for main stagesAny aspects not mentioned here remain the same as in the original model.