💡PonyXL Model Guide

What is PonyXL Model

Pony Diffusion V6 is a versatile SDXL fine-tuned and capable of producing stunning SFW and NSFW ( Note: Moescape AI does not support NSFW art generation. ) visuals of various anthro, feral, or humanoid species and their interactions based on simple natural language prompts. (Source: Pony Diffusion V6 XL - V6 (start with this one) | Stable Dif…)

Model Overall Review

🌟 Skill for NSFW & SFW image generation ( Note: Moescape AI do not support NSFW art generation. )

🌟 Offers various art styles (when Prompt & Lora applied)

🌟 Skill for multiple character generation

---

Pony Diffusion Image Gallery

Source

Source

Source

Source

Source

Source

---

Tips On Generating Better Images Using Pony Diffusion on Moescape AI

Tip 1: Add score tags in the prompt

Score Tags help the machine know what quality we want in the pictures. For example, using "score_9" in your prompt tells the machine to use the best-quality pictures it learned from.

Generation Comparison: With Score tags VS. without score tags

With Score Tags

Without Score Tags

Score Example:

• "score_9": Makes nice images but lacks noise. Perfection.

• "score_8_up": It has a lot of noise but little control. Chaos.

• "score_7_up": Has some noise and some control. Order.

• "score_6_up": It has some noise and control but is unstable. The Pretender.

• 'score_6": Similar to 6_up and is slightly more behaved. Little bother.

• "score_4_up" True middle ground. It contains more noise than "score_4" and is the vast populace of images. The Commoners.

• "score_4" Simmaler to "4_up". It is slightly more controlled but still has that unpredictable edge to it as "4_up"—the Commoners.

  Note: the score_4 tags are highly temperamental and will change depending on the seed and other settings used.

  "score_number_up" tags are generally more noisy, and "source" tags adhere to them more.

  Depending on the order of these score tags, you can lead your generation towards more chaotic and detailed or more controlled and straightforward.

What Score combos to use, and how to make your own:

For average, slightly smoother images:

[score_9,score_8_up,score_7_up]

For an on average, more detailed image, better backgrounds, or when you find an error:

[score_7_up,score_8_up,score_9]

For even more noise, use:

[score_8_up,score_7_up,score_9]

Score_# & Score_#_up Comparesion by @The Wanderer, Jax.

Score_#

Score_#_up

Tips By Moescape Discord Community Member @The Wanderer, Jax.

Some PonyXL models do better when you lead the "Score Tags" from lowest to highest. Ascending order.

Example:

• <Positive: score_7_up, score_8_up, score_9, (rest of your prompt).>

• <Negative: score_4, score_5, score_6, (Rest of negative prompt).>

List of tags that DO HAVE a large noticeable influence on generation:

• [(Score_number:Type tags.), "score_4", "score_5", "score_6/7", "score_8","score_9".]

• [(Score_number_up:Type tags.), "score_4_up", "score_6_up", "score_7_up", "score_8_up".]

Note: While the tag "score_9_up" has no real data behind it, it still offers minor desirable hallucination noise due to the "score_9" in its name. "score_9_up" can poison prompts and offer both minor enhancements or very noticeable corruptions (hands and fingers broken, animal ears missing, position flip) heavily based on seed and prompt it's used in. Experiment with its use, or preferably not at all.

Tags that DO NOT have any real influence on generation:

• [(Score_number), "score_1", "score_2", "score_3", "score_6".]

• [(Score_number_up, "score_1_up", "score_2_up", "score_3_up", "score_5_up(VERY SLIGHT change to face/lighting/shading)", "score_9_up(Negligible. Very temperamental. Read notes.)"

"score_8" is more chaotic and, on average, offers better results than "score_9" in clothing, background, and very subtle details.

---

What are score tags, & Why is it better to use them in your prompt?

(Simplified Explanation By official Pony Diffusion)

The Challenges

When the machine starts drawing, it doesn't know what "nice" means. If it learns from messy drawings, it will draw messy pictures. We want the machine to learn from good pictures, but finding enough good images is challenging, especially for specific things like ponies. Also, we can only sometimes tell which photos are good or bad.

Using the Rankings

Through our training, the machine has developed the ability to interpret tags like 'score_9' and 'score_8_up '. It now understands that 'score_9' signifies the best pictures, while 'score_8_up' indicates good pictures but not the best.

Making the Machine Draw Better

We add these tags to our prompts when we ask the machine to draw. This helps the machine know what quality we want in the pictures. For example, using "score_9" in your prompt tells the machine to use the best-quality pictures it learned from.

What Changed in V6?

In the latest version, V6, we wanted to ask for at least 80% good pictures. So, instead of just "score_8," we added tags like "score_8, score_7_up," which means "give me pictures that are at least 70% good.”

Do You Need to Use These Tags?

Maybe. Some tools automatically add these tags for you, while others require you to add them yourself. If you're using special styles or artist-specific settings, you might want to try without the tags to see how the machine reacts.

Note: These tags help make better pictures but don't work perfectly to avoid bad pictures. So, to avoid really bad images, you might need other methods.

Tip 2: Add Spells/LoRAs to achieve your desired art styles or concepts

Here are a few spells that can potentially improve the quality of your Pony XL model image generation

Nai Style For Pony

Generation Example:

Expressiveh

Generation Example:

“This is popular because the spell significantly improves the base PonyV6's understanding of 3D relationships in sexual positions and the resulting body shading. This makes sex scenes look more realistic. It also slightly enhances the level of sexual desperation in facial expressions.” (Credit: yokaikoibito478 )

“ExpressiveH is more than just a sexual LoRA. It is a detailer, prompt refiner, make things better-er, LoRA. The most ease of use LoRA we got. A side effect of this power is the staple "PonyXL" style you see growing on a lot of images, you can just tell it's PonyXL. Skin gets a little smoother and has this almost dreamlike hazy quality. Hate it or love it, it works GOOD.” (Credit: @the.wanderer.jax)

Cutesexyrobutts Style (Pony XL)

Generation Example:

Vixon's Pony(XL) Styles

Generation Example:

Granblue Fantasy PSXL | Pony Style Lora

(Granblue Fantasy PSXL | Pony Style Lora.) This is another beautiful LoRA mention. While not the best example in my plots, it still shows how the images feel more anime while not going deep into the 2d section. It "Flattens" out images and gives a nicer, more anime-like look than PonyXL and ExpressiveH skip.

Artist Style: Banpai Akira XL

Generation Example:

Nyantcha Style | Pony Diffusion

Ratatatat74 Artstyle | LoRA

Manhwa 3 in 1 PDXL & SDXL

Tip 3: Prompt order

Remember this: The order in which your prompt is drawn is from the first to the last tag, with each tag getting drawn each STEP. Tags at the start and the end have higher priority.

Example: [Most Priority, high priority, priority, medium, medium, medium, high medium.] The longer your prompt, the more the AI may forget your first tags.

---

Thank you for reading this guide. If you have more tips or questions, please share them on the PonyXL Discussion channel on our Discord!