【364】Pony V6 v6正式版

Pony V6 是一个多功能的 基础算法XL 微调模型，能够根据简单的自然语言提示生成各种拟人、野兽或人类物种及其互动的令人惊叹的 SFW 和 NSFW 视觉效果。
重要信息：

请确保以 clip skip 2（或某些软件中的 -2）加载此模型，否则你将获得低质量的图像。

此模型支持广泛的风格和美学，但提供了一个有见地的默认提示模板，允许在没有负面提示和默认设置的情况下生成高质量的样本：
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up，只需描述你想要的，tag1, tag2
（之前的 Pony Diffusion 模型使用了更简单的 score_9 质量修饰符，V6 XL 版本的较长版本是训练过程中无法及时纠正的训练问题，你仍然可以使用 score_9，但与完整字符串相比，其效果要弱得多。你可以在这里了解更多关于这些标签的信息）。

在大多数情况下，此模型不需要负面提示，也不需要其他质量修饰符，如 "hd"、"masterpiece" 等。

其他特殊的数据选择标签包括 'source_pony'、'source_furry'、'source_cartoon' 和 'source_anime' 以及 'rating_safe'、'rating_questionable' 和 'rating_explicit' 的评级。

此模型能够识别许多流行和不太知名的角色和系列。

如果你专门寻找小马风格，我建议使用以下两种模板之一 anthro/feral pony, rest of the prompt 或 source_pony, rest of the prompt。

此模型在自然语言提示和标签的组合上进行训练，并能够理解两者，因此在大多数情况下，使用正常语言描述预期结果是可行的，尽管你可以在主要提示后添加一些标签以增强它们。

建议使用 25 步的 Euler a 和 1024px 的分辨率，尽管模型通常可以处理大多数支持的 SDXL 分辨率。

此模型有时会生成难以用负面提示去除的伪签名，这是一个训练问题，将在未来的模型中纠正。如果这是你的问题，建议尝试 V5.5 或修复。

特别感谢

• Iceman 帮助采购必要的训练资源
• Haru 协助标注工作
• Cookie 在训练中的技术专长
• PSAI 服务器订阅者支持项目成本
• PSAI 服务器版主保持警惕并管理社区

技术细节
该模型基于作者个人偏好的美学排名，对约 260 万张图像进行训练，动漫/卡通/兽类/小马数据集的比例约为 1:1，安全/可疑/明确评级的比例约为 1:1。大约 50% 的所有图像都带有高质量的详细说明，这使得自然语言能力非常强。

所有图像均已使用说明（如有）和标签进行训练，艺术家的名字已被删除，并且根据我们的选择加入/退出计划对源数据进行了过滤。任何涉及未成年角色的明确内容已被过滤掉。

Pony Diffusion V6 is a versatile SDXL finetune capable of producing stunning SFW and NSFW visuals of various anthro, feral, or humanoids species and their interactions based on simple natural language prompts.

CHECK "ABOUT THIS VERSION" ON THE RIGHT IF YOU ARE NOT ON "V6" FOR IMPORTANT INFORMATION.

Please join our Discord Server to support development of new versions of this model and get access to free SD bot and check out more examples of this model capabilities on our prompt sharing website or follow the author on Twitter.

Important information

Make sure you load this model with clip skip 2 (or -2 in some software), otherwise you will be getting low quality blobs.

This model supports a wide array of styles and aesthetics but provides an opinionated default prompt template that allows generation of high quality samples with no negative prompt and otherwise default settings

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, just describe what you want, tag1, tag2

(previous Pony Diffusion models used a simpler score_9 quality modifier, the longer version of V6 XL version is a training issue that was too late to correct during training, you can still use score_9 but it has a much weaker effect compared to full string. You can learn more about these tags here).

The model is designed to not need negative prompts in most cases and does not need other quality modifiers like "hd", "masterpiece", etc...

Other special data selection tags include, 'source_pony', 'source_furry', 'source_cartoon' and 'source_anime' and ratings of 'rating_safe', 'rating_questionable' and 'rating_explicit'.

This model is capable of recognizing many popular and obscure characters and series.

If you are looking specifically for pony style, I recommend using one of the two following templates `anthro/feral pony, rest of the prompt` or `source_pony, rest of the prompt`.

This model is trained on combination of natural language prompts and tags and is capable of understanding both, so describing intended result using normal language works in most cases, although you can add some tags after the main prompt to boost them.

Using Euler a with 25 steps and resolution of 1024px is recommended although model generally can do most supported SDXL resolution.

This model will sometimes generate pseudo signatures that are hard to remove even with negative prompts, this is unfortunately a training issue that would be corrected in future models. If that's an issue for you I suggest trying V5.5 or inpainting.

Special thanks

• Iceman for helping to procure necessary training resources

• Haru for assistance with captioning efforts

• Cookie for technical expertise in training

• PSAI Server Subscribers for supporting the project costs

• PSAI Server Moderators for being vigilant and managing the community

Technical details

The model has been trained on ~2.6M images aesthetically ranked based on authors personal preferences, with roughly 1:1 ratio between anime/cartoon/furry/pony datasets and 1:1 ratio between safe/questionable/explicit ratings. About 50% of all images has been captioned with high quality detailed captions, which results in very strong natural language capabilities.

All images has been trained with both captions (when available) and tags, artists' names have been removed and source data has been filtered based on our Opt-in/Opt-out program. Any explicit content involving underage characters has been filtered out.

License

This model is licensed under a modified Fair AI Public License 1.0-SD (https://freedevproject.org/faipl-1.0-sd/) license.

The following modifications have been added to Fair AI Public License:

You are not permitted to run inference of this model on websites or applications allowing any form of monetization (paid inference, faster tiers, etc.). This applies to any derivative models or model merges.

If you want to use this model commercially, please reach us at contact@purplesmart.ai.

Explicit permission for commercial inference has been granted to CivitAi and Hugging Face.