PromptZone - AI Prompts, Guides and Tools for Builders

Tara Suzuki
Tara Suzuki

Posted on

Pony Diffusion V6 XL Prompt Guide: Tags and 12 Examples

Start Pony Diffusion prompts with score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, then describe the image in ordinary language and add relevant tags. For safe-for-work requests, include rating_safe; set Euler a, 25 steps, 1024px resolution, and clip skip 2 for V6 XL.

The settings and tag rules here come from PurpleSmartAI's V6 XL model page and score-tag explanation, published May 6, 2024. Start with the original V6 XL checkpoint; the Turbo variant has separate settings below.

What you want Start with Keep in mind
A stylized character The full score prefix and one clear sentence about the subject Get the basic character right before adding detail tags.
An anime or cartoon look source_anime or source_cartoon Leave the source tag out when you do not want that look.
A pony or furry character source_pony or source_furry Skip these tags for unrelated subjects.
A background or environment A description of the scene; see examples 3 and 10 Say that the scene is empty if nobody should be in it.
A Pony-based fine-tune That checkpoint's own instructions A name that contains "Pony" does not mean these settings carry over.

Pony Diffusion prompt tags: what each part does

Quality: copy the whole sequence

Use this exact quality prefix:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up
Enter fullscreen mode Exit fullscreen mode

These labels come from the author's aesthetic-ranking process. The intended _up meaning was a quality threshold and above, but V6 learned the full sequence as a signal for good-looking images instead of learning its components independently. The author describes this as a training issue discovered too late to correct: score_9 still works alone, but its effect is much weaker than the complete string.

Keep the sequence intact for your baseline. Its inclusion of score_4_up is therefore not an instruction to make a poor-quality image. The author also says V6 does not require extra quality words such as hd or masterpiece.

Source: choose the relevant visual category

The source tags are source_pony, source_furry, source_cartoon, and source_anime. The author calls these data-selection tags; use one when you want to request its corresponding category.

For pony-style images, the author recommends one of two openings: anthro/feral pony, rest of the prompt or source_pony, rest of the prompt.

Rating: keep these examples safe

Put rating_safe in the positive prompt for safe-for-work images. The model also has rating_questionable and rating_explicit; this guide does not use them.

Description: combine sentences and tags

V6 was trained with both natural-language captions and tags. The author's suggested arrangement is the quality prefix, a description of what you want, and additional tags after that description to emphasize details.

Write who or what is present, what it is doing, and where it belongs. Then specify framing, lighting, or medium. For a portrait, name the character’s occupation, clothing, and surroundings before choosing detail tags. Use ordinary words for those details; reserve the special tags for quality, source, and rating.

A copyable prompt skeleton

Use this working order: quality → description → source → rating → detail tags. Put the description before the trailing tags. Group source, rating, and detail tags in that order to keep this skeleton easy to edit.

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, [subject and action], [setting and composition], [medium and lighting], [optional source tag], rating_safe, [detail tags]
Enter fullscreen mode Exit fullscreen mode

Replace the bracketed fields before generation. Keep one central action in the first draft, and choose a framing that matches the job: a portrait, full-body pose, or wide scene.

For help articulating the scene, start with the image prompt generator, then organize your chosen description around this skeleton. Keep the exact V6 quality prefix while revising the creative parts. For prompting in Fooocus, see the Fooocus image prompt guide.

Get the daily AI digest
One email each morning — what shipped in AI, and the prompts we tested. No filler.

Pony Diffusion V6 XL settings

Setting V6 XL guidance
Sampler Euler a
Steps 25
Resolution 1024px; the description also says most supported SDXL resolutions generally work.
Clip skip 2, represented as -2 in some software
CFG The model page gives no number.
VAE The model page names none.
Negative prompt Usually unnecessary according to the author.

Set clip skip to 2, or -2 where your software uses that notation. Incorrect clip skip can yield low-quality blobs, so check it before rewriting the prompt.

Use 1024px as your starting resolution; the model page gives no table of aspect ratios, only that most SDXL resolutions generally work. It also gives no CFG value, so whatever your interface defaults to is your own choice, not the author's. When you compare two wordings of a prompt, keep every setting and the seed unchanged.

Choosing a Pony Diffusion negative prompt

An empty negative prompt is the author's baseline recommendation for most V6 requests. Leave the negative field blank to start, then add short exclusions only for things you want absent from your scene.

In the score-tag explanation, the author says quality tags are less useful in negatives because training only extends down to score 4. Low-score negatives cannot reliably steer away from the very worst images. Start without a long list of negative quality tags.

If you need ideas for exclusions, use the negative prompt generator selectively. Keep only terms that describe something you actually want absent, then compare against the empty-negative baseline.

12 complete Pony Diffusion prompt examples

These 12 examples are starting points written from the author's documentation; we did not run them. Every prompt uses rating_safe, and every character is an adult in everyday clothes. The negative prompts are optional: the model is designed not to need one in most cases, so start with the field empty and add one only if the thing it names shows up.

1. Character portrait

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, an adult botanist in a green jacket over a buttoned shirt, standing in a greenhouse, head-and-shoulders portrait, soft window light, leafy background, source_anime, rating_safe, glasses, calm expression
Enter fullscreen mode Exit fullscreen mode

Negative prompt: text, watermark

Adapt: Change the profession, jacket, and setting while keeping the portrait framing.

2. Full-body action pose

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, an adult mountain guide in a weatherproof jacket, trousers and hiking boots stepping across a stream, full body with boots visible, backpack, rocky valley, source_cartoon, rating_safe, side view
Enter fullscreen mode Exit fullscreen mode

Negative prompt: close-up, cropped feet

Adapt: Replace the action and terrain; keep the clothing and framing explicit.

3. Scenery or background

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, an empty mountain railway station at sunrise, wide establishing view, distant peaks, wooden platform and a red signal lamp, painted animation background, rating_safe, mist, warm light
Enter fullscreen mode Exit fullscreen mode

Negative prompt: people, animals, vehicles

Adapt: Swap the location and time of day while retaining the empty-scene brief.

4. Animal illustration

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, a red fox standing beside an old oak tree, autumn woodland, full animal visible, soft painted fur, fallen leaves, source_cartoon, rating_safe, side view, morning light
Enter fullscreen mode Exit fullscreen mode

Negative prompt: clothing, accessories

Adapt: Replace the animal and habitat together to keep the scene coherent.

5. Fantasy creature concept

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, a moss-covered stone dragon resting beside a waterfall, broad rounded wings folded against its back, full creature visible, fantasy concept painting, rating_safe, ferns, overcast light
Enter fullscreen mode Exit fullscreen mode

Negative prompt: text, watermark

Adapt: Change the creature's material, silhouette, and environment as a coordinated set.

6. Art Nouveau poster illustration

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, an adult violinist in a long high-collared coat, Art Nouveau poster illustration, flowing botanical curves, flat muted colors, ornamental floral border, centered composition, rating_safe, decorative linework
Enter fullscreen mode Exit fullscreen mode

Negative prompt: lettering, typography, watermark

Adapt: Replace the instrument and flowers while retaining the named style and visual description.

7. Manga panel

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, one manga panel showing an adult detective in a belted raincoat and hat opening an office door, black-and-white ink, diagonal shadows, medium shot, source_anime, rating_safe, screentone shading
Enter fullscreen mode Exit fullscreen mode

Negative prompt: speech bubbles, lettering, multiple panels

Adapt: Change the action and setting, keeping the single-panel framing.

8. Pony character

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, a four-legged chestnut pony standing beside a wooden orchard fence, braided mane, apple trees and distant hills, full-body side view, source_pony, rating_safe, afternoon sunlight
Enter fullscreen mode Exit fullscreen mode

Negative prompt: none; leave the field empty.

Adapt: Change the coat, mane, and surroundings to define a different character brief.

9. Anthropomorphic explorer

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, an adult anthropomorphic otter explorer in a canvas jacket, trousers and boots, holding a compass on a riverside trail, three-quarter full-body view, source_furry, rating_safe, satchel, relaxed stance
Enter fullscreen mode Exit fullscreen mode

Negative prompt: text, watermark

Adapt: Replace the species and exploration equipment without changing the basic pose.

10. Interior environment

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, an empty observatory library with tall bookshelves and a brass telescope, wide interior view, circular window, blue evening light, painted background, rating_safe, wooden floor, quiet atmosphere
Enter fullscreen mode Exit fullscreen mode

Negative prompt: people, animals

Adapt: Change the room's purpose, central object, and light source.

11. Game prop concept

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, a single mechanical compass with a brass case and blue enamel lid, isolated game prop concept, three-quarter view, neutral gray background, painted metal textures, rating_safe, open lid
Enter fullscreen mode Exit fullscreen mode

Negative prompt: hands, people, lettering

Adapt: Replace the object and materials while keeping the isolated presentation.

12. Editorial teamwork illustration

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, two adult engineers in work jackets and trousers assembling a small tabletop robot, one holds a screwdriver and the other reads a notebook, medium-wide workshop view, source_cartoon, rating_safe, bright colors
Enter fullscreen mode Exit fullscreen mode

Negative prompt: lettering, watermark

Adapt: Assign each adult one distinct task and replace the shared project.

Troubleshooting V6 prompts

Symptom Likely cause Fix to try
Low-quality blobs Incorrect clip skip. Set clip skip to 2, or the software's equivalent -2.
Weak quality guidance with score_9 alone V6 learned the longer sequence as a quality signal. Restore the complete six-tag prefix.
Signature-like marks persist A training issue leaves pseudo-signatures that negatives may not remove. Try inpainting.
A style LoRA reacts poorly to the score prefix Score tags carry biases; compare their effect with your style LoRA. Compare with the score tags omitted while retaining rating_safe.
Low-score negatives do little The score-tag training range does not reach the lowest-quality examples. Return to the empty-negative baseline instead of adding more score exclusions.

If a LoRA is involved, consult the ComfyUI LoRA loading and troubleshooting guide before changing several prompt components at once.

Pony-based fine-tunes and later versions

For Pony-based fine-tunes, follow the exact checkpoint's documentation. If you are choosing another SDXL checkpoint, consult the SDXL model guide before borrowing this prompt template. The author's model listing includes V6 Turbo DPO merge, published January 14, 2024: a merge by Zatochu combining Pony Diffusion with SDXL Turbo and DPO. Its own instructions are Euler A at 10 to 15 steps with a scale of 2 to 3, and they warn that the merge may lose some of the model's understanding of characters and composition. Those values are for the Turbo merge only, not for base V6 XL.

Pony V7 is built on a different architecture. Its model card says it is based on AuraFlow, and that special tags, quality tags included, have a much weaker effect than in V6, so score_9 will not necessarily improve a prompt. The card gives V7 its own prompt structure; follow that instead of this guide's skeleton. The V7 weights were uploaded on October 23, 2025, according to the repository history.

FAQ

Do Pony Diffusion prompts need the full score string?

For V6 XL, use the complete six-tag sequence as your baseline. The author says score_9 alone remains usable but has a much weaker effect because of how V6 learned the training labels.

What is the best Pony Diffusion negative prompt?

The author's baseline is usually an empty negative prompt. Start there, then compare short exclusions tied to your particular brief, such as objects or text you want absent.

Can I use natural language instead of only tags?

Yes: V6's author documents training with both captions and tags. Describe the intended scene normally, then append relevant tags after the description.

What CFG and VAE should I use for Pony Diffusion V6 XL?

The author's model page gives no CFG number and names no VAE for V6 XL. What it does recommend is Euler a, 25 steps, 1024px and clip skip 2. The scale of 2 to 3 you may see quoted belongs to the Turbo merge, not to the base model.

Get the daily AI digest

One email each morning — what shipped in AI, and the prompts we tested. No filler.

Top comments (0)