Pony Diffusion V6 XL

Pony Diffusion V6 XL represents a significant advancement in specialized image generation, built on the Stable Diffusion XL architecture. This model focuses on generating high-quality images of anthropomorphic, feral, and humanoid characters, with particular emphasis on pony-themed content.

Training and Data

The model's strength lies in its extensive training dataset of approximately 2.6 million aesthetically ranked images, with roughly half containing detailed captions. This careful curation ensures strong natural language understanding and high-quality output generation. The training data represents a balanced distribution across multiple artistic styles:

• Anime
• Cartoon
• Furry
• Pony

The dataset maintains an even distribution across safe, questionable, and explicit ratings, with specific exclusions for inappropriate content. The creators implemented an opt-in/opt-out program for artists, ensuring ethical data sourcing while removing artist names from the training process.

Technical Specifications and Usage

For optimal results, the model performs best with the following settings:

• Euler a sampler with 25 steps
• 1024px resolution (though most SDXL resolutions are supported)
• Clip skip of 2 (or -2 in some software)

The model currently generates pseudo-signatures that can be difficult to remove, an issue scheduled for correction in future versions. For specific pony-style generations, users should structure prompts as:

anthro/feral pony, [rest of prompt]

or

source_pony, [rest of prompt]

Model Variants and Features

Pony Diffusion V6 XL comes in several variants:

1. V6 (recommended base version)
2. V6 Turbo DPO merge
3. V6 Turbo merge
4. V6-1.5

The V6 version serves as the recommended starting point for most users. One notable feature in V6 XL is its modified prompt template structure:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, [description], tag1, tag2

While this longer template results from a training artifact, the simpler score_9 quality modifier from previous versions remains functional, albeit with reduced effectiveness.

Licensing and Distribution

The model operates under a modified CreativeML Open RAIL++-M license with specific restrictions on monetized inference. Commercial use requires direct contact with the creators, though exceptions have been granted to specific platforms like CivitAi and Hugging Face. For detailed licensing information, refer to the Fair AI Public License 1.0-SD.

References

1. Civitai Model Page - Primary model distribution and documentation
2. Pony Diffusion Discord - Community support and updates
3. PurpleSmartAI - Example outputs and prompts
4. Usage Tips - Detailed guide for optimal results
5. Artist Opt-in/Opt-out - Data sourcing program information
6. Model Tutorial - Video guide for model usage
7. Recommended VAE - Optional enhancement for output quality