Collective Recipe

使用方法 | Useage：

trigger word:

stylized

模型简介 | Intro：

这是一组由AI生成图片训练而成的画风LoCon。

This is a set of style LoCon trained on AI generated images.

Pony Diffusion V6 是 SDXL在开源初期最成功的几个微调项目之一。

随着时间推移，新的训练技术的开发、新的数据集的应用使得近来发布的开源模型在性能上明显超过了Pony V6，基于Pony V6的模型在Civit上活跃度也不如以前。

但与大多数基于Danbooru训练的二次元模型不同，Pony V6的独特生态给AI社区带来了许多前所未有的艺术风格。即使新的开源模型知识更丰富、生成更稳定，依然不能完全代替pony的生态。

这个系列的模型旨在对过去Civit社区曾经流行过的一些艺术风格组合进行复刻，把这些经典的美术风格迁移到更新、更稳定的先进模型上。这些风格是从我收藏的图片中选定的。每一套风格都会锚定一组固定的Lora、prompt参数。

出于对数据集质量、结构和一致性的诸多考虑，整个训练不会使用其他用户已经发布的图片，而是使用工作流重新生成、人工筛选。我设计了一套提示词组合，希望可以尽量均衡地从模型中提取不同主题的画面内容，保证模型能适应足够多的任务。

Pony Diffusion V6 was one of the most successful SDXL fine-tuning projects during the early days of its open-source release.

Over time, advancements in training techniques and the introduction of new datasets have led to the release of open-source models that significantly outperform Pony V6 in terms of generation quality. Models based on Pony V6 have also seen a decline in activity on Civit.

However, unlike most anime-style models trained on Danbooru, Pony V6 fostered a unique ecosystem that introduced a wide range of unprecedented art styles to the AI community. Even though newer models offer more knowledge and greater stability, they still cannot fully replace the creative diversity found within the Pony ecosystem.

This model series aims to replicate some of the art style combinations that were once popular in the Civit community, transferring those classic art styles onto newer, more stable and capable models. These styles were selected based on images from my personal collection. Each style in the series is anchored by a fixed set of LoRAs and prompt parameters.

Due to various considerations regarding dataset quality, structure, and consistency, no previously published user images will be used in training. Instead, all data will be regenerated via workflows and manually curated. I have designed a prompt structure intended to extract a balanced range of visual themes from each model, ensuring that the model remains adaptable to a wide variety of tasks.

因为Civit现在的生成服务涉及到模型制作者和用户的经济利益，我会优先选取那些涉及的模型资源多、在原有模型架构上可控性低的组合；跳过那些只有2-3个资源的简单组合。那些比较新的模型，尤其是使用了新的checkpoint的组合我也会尽量避开。总之，是尽量少地直接损害还在活跃的Pony模型作者利益，也给想要生成一些经典画风的生成用户提供一个更经济、更为稳定可控的生成方案。

Since Civit’s current generation services involve economic interests between model creators and users, I will prioritize combinations that rely on more resources and exhibit low controllability. Simpler combinations based on only 2–3 resources will be skipped. I also intend to avoid recent models, especially those generated on new checkpoints. In short, this project tries to minimize any direct harm to active Pony model creators, while offering users a more economical and stable alternative to generate classic artistic styles.

配方 | Recipe：

Recipe1:

Checkpoint: Pony Diffusion v6 XL

Other Resources:

All Disney Princess XL LoRA Model from Ralph Breaks the Internet v2.0 (weight:0.7)

ExpressiveH (Hentai LoRa Style) エロアニメ (weight:0.4)

Incase Style [PonyXL] v3.0 (PonyXL Edition) (weight:0.4)

Vixon's Pony Styles - gothic neon v1.0 (weight:0.2)

Prompt:

score_9, score_8_up, score_7_up, score_6_up, source_anime, source_comic, flat color, Expressiveh,