Pony & Illustrious Prompt Syntax: Score, Source, Rating Tags

TL;DR

• Score tags differ between V6 and V7: V6 uses a 6-tier system (score_9 through score_4_up) mapped to training-data percentiles; V7 weakens these tags — use score_8 or score_7_up instead of score_9 for better results. Source
• Source and rating tags only work if the model was trained on them: source_pony, source_furry, source_anime, source_cartoon and rating_safe/questionable/explicit require the training dataset to include those labels — they don't transfer to merges or fine-tunes automatically. Source
• CLIP skip 2 is mandatory for Pony models: Loading Pony Diffusion without clip_skip=2 (or -2 in some UIs) produces low-quality blobs regardless of prompt quality. Source

Why the Official Score Tag Advice Is Misleading

The Civitai article introducing score_9 presents it as a magic quality button: "put score_9 in your prompt, get the best images." That's not what happens in practice, and the model creators later clarified why.

Pony Diffusion V6 was trained with a six-tier quality labeling system applied to the training images themselves. Each image received a score_4 through score_9 tag based on human quality assessment. The _up variants (score_8_up, score_7_up, etc.) were added later to let users request cumulative ranges — for example, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up means "anything from the 40th percentile up." Source

The correction that matters: In V6, score_9 targets the top 10% of training data. In V7, PurpleSmartAI explicitly states 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." Community testing confirms score_8 or score_7_up often produce more coherent clothing, backgrounds, and fine details in V7. Source

Related: How CLIP skip affects prompt adherence in SDXL models

Source_ and Rating_ Tags: Dataset Labels, Not Magic Switches

Four source_ tags exist in Pony V6: source_pony, source_furry, source_anime, source_cartoon. Three rating_ tags: rating_safe, rating_questionable, rating_explicit.

Critical correction from the source material: "If a model trainer uses these tags to describe images in the dataset they train on, these tags will work. If they don't, these tags will influence the resulting image, but only because it's trying to understand what you mean by 'score_9'." Source

This means:

• They are not style transfers. Adding source_anime doesn't "make it anime style" unless the model learned that association during training.
• They don't work on merges that dropped the labeled data. A Pony/Anime merge that didn't preserve the source tags will ignore them.
• Negatives behave differently. The Civitai guide notes: "Put it in the negative, and it will try to avoid it" — but this is model-dependent.

Illustrious XL documentation mentions "natural language + tag-based prompting" but does not confirm it uses the same source_/rating_ taxonomy. Treat them as Pony-specific unless the model card says otherwise. Source

The Mandatory CLIP Skip 2 Setting

Every Pony Diffusion model page carries the same bold warning: "Make sure you load this model with clip skip 2 (or -2 in some software), otherwise you will be getting low quality blobs." Source This is not optional. CLIP skip changes which transformer layer's output feeds the U-Net. Pony models were trained with CLIP skip 2 baked in — the text encoder expects the penultimate layer, not the final layer. Without it, token embeddings drift and the model cannot align prompts to images. Source

| ComfyUI | CLIPTextEncode node: clip_skip: 2 | Civitai |

| InvokeAI | Model config: clip_skip: 2 | Civitai |

| Stable Diffusion CLI | --clip_skip 2 or -2 | Civitai |

Prompt Template: The Official V6 Structure

The model card prescribes this exact ordering:

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

Why this order matters: CLIP processes tokens sequentially. Quality tags first establish the "quality manifold" before subject tags refine within it. Danbooru-style tags (tag1, tag2) come last as fine-grained steering. Source

For V7, drop score_9 and lead with score_8, score_7_up or just score_7_up, score_6_up. Source

Weight Syntax: Explicit and Implicit

Two weight systems coexist in AUTOMATIC1111 and compatible UIs: Source

Explicit: (keyword: factor)

• Factor > 1 increases attention, < 1 decreases
• Practical range: 0.7–1.5 (outside this range distorts images) Source
• Example: (masterpiece: 1.2), (low quality: 0.7) Source

Implicit: Parentheses and brackets (multiplicative)

Combined: [((word))] = 0.9 × 1.21 ≈ 1.089

Delimiters: Comma, Dot, BREAK

BREAK must be uppercase. It forces the 75-token CLIP chunk boundary, preventing token bleed between unrelated concepts. Source

Prompt Scheduling (Blending)

Format: [keyword1: keyword2: factor] where factor (0–1) sets the step percentage to switch. Source

[daytime: nighttime: 0.5]

At 30 steps: steps 1–15 use "daytime", steps 16–30 use "nighttime". Creates transitional blending.

Confirmed working in: AUTOMATIC1111 WebUI. Not confirmed in ComfyUI, InvokeAI, or other backends.

Negative Prompts: What Actually Works

Score tags in negatives: limited utility

"You will likely also see score_ tags used in negatives, but they aren't as effective. Some prompters will go as high as score_6, but generally, you'd be okay with a negative prompt of 'score_5, score_4'." — Civitai guide Source

The doesnotexist trap

Critical: Never use doesnotexist without weighting. The Moescape guide explicitly warns: "Don't use the doesnotexist tag without weakening as here - (doesnotexist:0.65) otherwise it causes errors and artifacts." Source

Community-sourced negative tag pool

Common anti-artifact tags (community-reported, verify per model): Source

bad quality, bad anatomy, low quality, low res, blurry, jpeg artifacts,
worst quality, ugly, deformed, extra limbs, missing limbs, floating limbs,
disconnected limbs, mutation, mutated, poorly drawn face, poorly drawn hands,
bad proportions, gross proportions, malformed, missing fingers,
extra digits, fewer digits, cropped, out of frame, watermark, signature,
text, error, username, blurry, bad feet, bad hands

Related: Building effective negative prompts for SDXL

Background Tags That Work in Pony / Illustrious

Pony is "way more picky at backgrounds than SD1.5" — random location tags often produce non-Euclidean geometry. NanashiAnon's tested list (community-reported, works in both Pony and Illustrious): Source

Tip: Use fewer, more specific tags. Pony blends poorly; Illustrious handles backgrounds better.

Illustrious XL: Different Model, Different Rules

Illustrious XL (OnomaAIResearch) is not a Pony fine-tune — it's a separate SDXL-based train with these documented differences: Source

Do not assume score_9 or source_anime work in Illustrious. The model card only confirms natural language + tag prompting and 1536×1536 native resolution.

Related: Illustrious XL vs Pony Diffusion comparison

Common Errors and Fixes

FAQ

What is the correct score_ tag order for Pony V6 XL?

Use score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up at the start of your prompt. This mirrors the training data's cumulative quality buckets (40–100%). For V7, lead with score_8, score_7_up instead. Source

Do source_ tags work in negative prompts?

The Civitai guide says "put it in the negative, and it will try to avoid it," but this only works if the model was trained on those labels. Test per model; don't assume transfer. Source

Why does Pony V7 ignore score_9?

PurpleSmartAI states V7's "special tags (including quality tags) have much weaker performance compared to V6." The model relies more on prompt adherence than quality-tag conditioning. Source

What CLIP skip does Illustrious XL need?

The model card doesn't specify. Most SDXL derivatives use CLIP skip 1. Start with 1; only change if outputs degrade. Source

Can I use BREAK in ComfyUI?

Not officially confirmed. BREAK is an AUTOMATIC1111 WebUI feature that manipulates CLIP chunk padding. ComfyUI handles conditioning differently — test before relying on it. Source

What's the difference between (word:1.2) and ((word))?

(word:1.2) applies a flat 1.2× multiplier. ((word)) applies 1.1×1.1 = 1.21×. They're close but not identical; explicit weights give precise control. Source

Do score_ tags work in Illustrious XL?

No documented evidence. Illustrious uses natural language + Danbooru tags without a published quality-tier system. Source

How do I write a proper negative prompt for Pony?

Start with (doesnotexist:0.65), score_5, score_4, then add structural anti-tags: bad anatomy, bad hands, extra limbs, missing limbs, blurry, low quality, watermark, text. Avoid overloading — too many negatives conflict. Source, Moescape

Sources

1. Civitai: What is score_9 and how to use it in Pony Diffusion — Score tag system design, V6/V7 differences, negative prompt guidance, weight syntax, blending, delimiter behavior, background tags
2. Civitai: Pony Diffusion V6 XL model page — Official prompt template, CLIP skip 2 requirement, source_/rating_ tag list
3. Hugging Face: PurpleSmartAI Pony V7 Base — V7 quality tag weakening statement
4. Moescape: PonyXL Model Guide — doesnotexist weighting warning, score_8 vs score_9 comparison
5. Hugging Face: OnomaAIResearch Illustrious XL v1.0 — 1536×1536 native resolution, natural language + tag prompting, LoRA/ControlNet compatibility, dark/light generation
6. AUTOMATIC1111 Wiki: Features — Attention/emphasis syntax, BREAK keyword, infinite prompt length chunking