Pony Diffusion score_9 and score_8_up quality tags explained

TL;DR

• Start every Pony V6 prompt with the full chain score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up. A lone score_9 is too weak to do the job.
• Add a source_* tag (source_pony, source_furry, source_anime, source_cartoon) and a rating_* tag (rating_safe, rating_questionable, rating_explicit) to control style and content safety.
• Do not use Pony score tags on Illustrious or NoobAI-XL. Those models expect masterpiece, best quality and a year modifier such as newest. In Pony V7, score tags are explicitly weaker than in V6.

What does score_9 actually do in Pony Diffusion?

score_9 is a quality-conditioning token. Pony V6's creator manually ranked the training images and prefixed each caption with a score tag, so score_9 labels the top-ranked images and lower numbers weaker ones. Prepending the tag steers the model toward that top-ranked training distribution, much as generic SD models learn masterpiece from caption corpora — except Pony's scores come from a manual 1–9 ranking rather than caption frequency.

How the score/source/rating prompt syntax works

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

(Pony Diffusion V6 XL model card)

The model card states the V6 defaults: Euler a, 25 steps, ~1024px, and CLIP skip 2 — without clip skip 2 you get "low quality blobs" (Pony V6 card). In ComfyUI the same setting is -2 because the UI signs clip skip opposite to A1111 (Civitai workflow guide). For more on the sampler/step side of these defaults, see Stable Diffusion samplers and schedulers explained.

Why the long score_9, score_8_up chain is fused

The _up suffix is a historical accident. Earlier Pony versions used a simple score_9 token, but V6 expanded each label into the verbose chain — too late in training to correct. The model learned the whole string as the correlate of "good image" rather than each part independently, so the tags do not compose the way the _up naming suggests.

The card documents the consequence: "you can still use score_9 but it has a much weaker effect compared to full string" (Pony V6 card). A guide adds: "Using only score_9 usually don't do too well. You need at least 3," and notes the _up suffix is cosmetic (stable-diffusion-art). For the negative prompt, the same guide notes practitioners "sometimes put 'score_5_up, score_4_up' in the negative prompt" (stable-diffusion-art). The scale stops at score_4, so negatives only nudge away from mediocre images, not rescue a broken generation.

Do Illustrious models understand Pony score_9 tags?

No. Illustrious XL is an SDXL anime model trained on Danbooru tags, and its quality vocabulary is a set of percentile words, not Pony's score_9 token. An Illustrious guide states: "When prompting illustrious, you do not, run score tags like you do with pony as the model won't understand them" (Arctenox Illustrious guide).

Illustrious expects masterpiece, best quality in the positive prompt and worst quality, low quality in the negative. Its technical paper buckets quality by Danbooru score percentile into worst quality (~8%) up to masterpiece (~100%), and adopts a NovelAI tag-ordering schema (person count, character, rating, general tags, artist, score-range quality, year) (Illustrious arXiv paper). For a deeper anime prompting walkthrough, see Illustrious XL prompting for anime characters.

Illustrious and NoobAI quality tag systems

Illustrious pairs quality words with year/period modifiers; NoobAI-XL, a v-prediction finetune of Illustrious v0.1 (Danbooru + e621), inherits the same approach with its own cutoffs.

Illustrious v2.0 supports 512–1536px with stronger natural-language handling (Illustrious v2.0 card).

The two ecosystems also differ on content safety: Illustrious v2.0 exposes explicit, questionable, sensitive, general, plus alternates like nsfw and sfw, whereas Pony uses rating_safe / rating_questionable / rating_explicit (Illustrious v2.0 card).

Pony V7 changed everything: score tags are weaker now

Pony V7 is a rebuild on AuraFlow, a 7 billion parameter architecture, trained on roughly 10 million images from a superset of over 30 million, 100% captioned — a deliberate break from V6's SDXL base and ~2.6 million-image dataset (Pony V7 card). For the architecture story, see Pony V7 on AuraFlow versus SDXL V6.

V7 retains the score_X, style_cluster_x, and source_X special tags, but the card warns "V7 prompting may be inconsistent" and that "Special tags (including quality tags) have much weaker performance compared to V6, meaning score_9 would not necessarily yield better results on some prompts." Its template favors natural language: "special tags, factual description of image, stylistic description of image, additional content tags" (Pony V7 card).

When quality tags help or hurt

Quality tags help when you want the model to commit to its top-tier training distribution — most of the time on Pony V6, which is designed to need no negative prompt or extra modifiers like "hd" or "masterpiece" (Pony V6 card). They hurt — or simply do nothing — in three situations:

1. On the wrong model. score_9 in an Illustrious or NoobAI prompt wastes tokens — those models were never trained on it. Use masterpiece, best quality instead (Arctenox guide).
2. On Pony V7. Score tags are weaker than in V6, so relying on them as in V6 gives inconsistent results. Lead with natural-language description (Pony V7 card).
3. In the negative prompt at the wrong end. Pony's scale stops at score_4, so negatives only nudge away from mediocre images. Community practice puts score_5_up, score_4_up in the negative — a mild push, not a rescue (stable-diffusion-art).

Common errors and fixes

FAQ

What does score_9 actually do in Pony Diffusion?

score_9 is a quality-conditioning token. Pony V6's creator manually ranked training images and prefixed each caption with a score tag; score_9 labels the best, and prepending it steers the model toward that top-ranked distribution. (Pony V6 card)

Do I need the full score_9, score_8_up, score_7_up chain, or is score_9 alone enough?

score_9 alone is too weak. The V6 card says the full chain has a much stronger effect, and a guide says "you need at least 3" tags. Use the full six-tag chain, or at minimum the top three. (Pony V6 card; stable-diffusion-art)

What is the difference between Pony score tags and Illustrious masterpiece/best quality tags?

Pony uses numeric score tokens (score_9, score_8_up) from a manual 1–9 ranking. Illustrious uses Danbooru percentile words (masterpiece ~100% down to worst quality ~8%) plus year modifiers. They are not interchangeable because each model was trained on a different vocabulary. (Illustrious arXiv)

Do Illustrious models understand Pony's score_9 tag?

No. An Illustrious guide says the model "won't understand them." Use masterpiece, best quality (positive) and worst quality, low quality (negative), with a year modifier such as newest. (Arctenox guide)

What do source_pony, source_anime, source_furry, and source_cartoon do?

They are Pony V6's domain tags. Each conditions the model toward a visual style — pony, anime, furry, or cartoon — by steering toward that domain's training subset. Place them after the score tags; they also work in the negative prompt. (stable-diffusion-art)

What is the difference between rating_safe, rating_questionable, and rating_explicit?

These are Pony V6's content-safety tags. rating_safe produces SFW output, rating_questionable sits in between, rating_explicit allows NSFW. Use them in either prompt to request or suppress a level. (stable-diffusion-art)

Why does Pony V6 need CLIP skip 2?

The V6 card warns that without clip skip 2 you get "low quality blobs." Pony V6 was trained against a specific layer of the SDXL text encoder, and clip skip 2 selects it. In ComfyUI the value is -2. (Pony V6 card)

Are score tags still useful in Pony V7?

Less than in V6. The V7 card says special tags including quality tags "have much weaker performance compared to V6," and flags V7 prompting as inconsistent while V7.1 is in progress. V7 favors natural-language description. (Pony V7 card)

Can I put score_5_up and score_4_up in the negative prompt?

Yes, as a mild nudge. The Pony scale stops at score_4, so negatives can only push away from mediocre images, not genuinely broken ones. Community practice puts score_5_up, score_4_up in the negative. (stable-diffusion-art)

What quality tags does NoobAI-XL use, and how do they differ from Pony?

NoobAI-XL uses percentile words: masterpiece (>95th), best quality (85th–95th), good quality (60th–85th), normal quality (30th–60th), worst quality (≤30th), plus period tags like newest. Prefix: masterpiece, best quality, newest, absurdres, highres, safe,. It is a v-prediction model needing Euler and a dynamic CFG plugin. (NoobAI-XL card)

Sources

• Pony Diffusion V6 XL model card — https://civitai.com/models/257749/pony-diffusion-v6-xl
• Pony V7 base model card — https://huggingface.co/purplesmartai/pony-v7-base
• Illustrious XL v0.1 model card — https://civitai.com/models/795765/illustrious-xl
• Illustrious XL v2.0 model card — https://civitai.com/models/1369089/illustrious-xl-20
• Illustrious technical paper (arXiv) — https://arxiv.org/html/2409.19946v1
• NoobAI-XL model card — https://civitai.com/models/833294/noobai-xl-nai-xl
• Pony Diffusion prompt tags guide — https://stable-diffusion-art.com/pony-diffusion-prompt-tags/
• Arctenox Illustrious prompt guide — https://civitai.com/articles/23210/arctenoxs-simple-prompt-guide-for-illustrious-for-comfyuisite
• CLIP ViT-L/14 model card — https://huggingface.co/openai/clip-vit-large-patch14
• Pony V6 ComfyUI workflow guide — https://civitai.com/articles/4975