Pony Diffusion V6 XL — Versatile SDXL Fine-Tune

Pony Diffusion V6 is a versatile SDXL fine-tuned model designed to generate a broad range of high-quality SFW and NSFW imagery, featuring anthro, feral, and humanoid characters, diverse species, and their interactions. It supports both simple natural-language descriptions and detailed tag-based prompting, allowing users to explore a wide variety of artistic styles and visual concepts.

IMPORTANT: Make sure you are using the correct V6 version. Check the “About This Version” section on the right side of the model page to confirm that V6 is selected and review any version-specific information.

Join our Discord community to support the development of future model versions, access the free SD bot, discover additional examples through our prompt-sharing website, and follow the author's updates on Twitter.

Important Information

Recommended Model Configuration

For optimal results, load the model with CLIP Skip set to 2 (or -2 in software that uses an alternative convention). Incorrect CLIP Skip settings may result in degraded image quality, including poorly defined or blob-like outputs.

Pony Diffusion V6 supports a broad spectrum of artistic styles and visual aesthetics. Its training incorporates a recommended prompt structure that can produce high-quality results without requiring a negative prompt or additional quality modifiers.

Default Prompt Structure

Start with the following quality-tag sequence, followed by a natural-language description of the desired image and any additional relevant tags:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, [your prompt], [additional tags]

Earlier Pony Diffusion models used the simpler score_9 quality modifier. In V6 XL, the longer sequence reflects a training-related limitation that could not be corrected before training was completed. Although score_9 remains usable on its own, its effect is considerably weaker than that of the complete sequence.

Further information about these quality tags is available through the model's associated documentation.

Negative Prompts and Quality Modifiers

The model is designed to work effectively without a negative prompt in most situations. Additional terms such as masterpiece, HD, or similar generic quality modifiers are generally unnecessary and may not provide meaningful improvements.

For greater control over image composition, style, or subject characteristics, focus on describing the intended result clearly and add specific tags only when needed.

Special Source and Rating Tags

The model recognizes several specialized tags that help guide the type of content or artistic domain:

• source_pony — pony-oriented visual styles.
• source_furry — furry and anthropomorphic characters.
• source_cartoon — cartoon aesthetics.
• source_anime — anime-inspired imagery.

Content-rating tags include:

• rating_safe
• rating_questionable
• rating_explicit

These tags can be combined with natural-language descriptions to guide generation toward the intended style or rating category.

Character Recognition and Pony-Specific Prompting

Pony Diffusion V6 can recognize numerous popular and less widely known fictional characters, franchises, and series.

For pony-focused imagery, the following prompt structures are recommended:

anthro/feral pony, [your prompt]

or

source_pony, [your prompt]

Both approaches provide a useful starting point for generating pony-style characters and scenes.

Natural Language and Tag-Based Prompting

The model was trained using a combination of natural-language captions and conventional tags. As a result, it can interpret ordinary descriptive sentences as well as structured tag sequences.

In most cases, simply describing the desired subject, appearance, pose, environment, lighting, and interaction is sufficient. Additional tags can be appended to emphasize specific visual attributes or reinforce particular concepts.

Recommended Generation Settings

The following settings provide a practical starting point for image generation:

• Sampler: Euler a
• Sampling steps: 25
• Resolution: 1024 px, with compatible SDXL resolutions also supported
• CLIP Skip: 2

Although these settings are recommended, the model can operate across most supported SDXL resolutions. Different compositions and artistic styles may benefit from individual adjustments.

Known Limitation: Pseudo-Signatures

Some generations may contain artificial or pseudo-signature markings that are difficult to eliminate, even when negative prompts are used. This is a known limitation associated with the training data.

If these artifacts interfere with your workflow, consider using Pony Diffusion V5.5 or removing the unwanted markings through inpainting.

Special Thanks

Sincere thanks to everyone who contributed to the development of this model:

• Iceman — assistance in obtaining essential training resources.
• Haru — support with image captioning.
• Cookie — technical expertise and training assistance.
• PSAI Server Subscribers — financial support for project expenses.
• PSAI Server Moderators — community management, moderation, and ongoing support.

Technical Details

Pony Diffusion V6 was trained on approximately 2.6 million images, aesthetically ranked according to the author's personal preferences.

The training dataset was organized with approximately:

• A 1:1 distribution between the anime, cartoon, furry, and pony dataset groupings.
• A 1:1 distribution between safe, questionable, and explicit content ratings.
• High-quality, detailed captions for approximately 50% of the images, strengthening the model's natural-language comprehension and descriptive prompting capabilities.

All images were trained using captions when available, alongside associated tags. Artist names were removed, and source data was filtered according to the project's opt-in/opt-out program.

The training dataset was also filtered to exclude explicit content involving underage characters.

License and Usage Restrictions

This model is distributed under a modified version of the Fair AI Public License 1.0-SD:

https://freedevproject.org/faipl-1.0-sd/

The model includes additional restrictions beyond the underlying license.

Commercial inference restriction: You may not run inference using this model on websites or applications that permit any form of monetization, including paid inference or paid faster-generation tiers. This restriction also applies to derivative models and model merges.

For commercial use, contact:

contact@purplesmart.ai

Explicit permission for commercial inference has been granted to Civitai and Hugging Face, as specified by the model's license terms.

Please review the complete license and its modifications before deploying the model, distributing derivatives, or using it in a commercial service.