PixAI SDXL Prompting Guide: Essential Tips & Tricks

Editor’s note (PixAI)

This guide was written by one of our community’s standout creators, 阿童 (ATone), and is republished here with credit to the original author.

Original website
ATone on PixAI

This chapter covers the fundamentals of prompt engineering, perfect for newcomers looking to master PixAI quickly.

Table of Contents

PixAI Core Features Breakdown

Prompting Tools

Feature	Description
Prompt Helper	An integrated LLM that rewrites your input into optimized prompts before generation. Supports natural language (and Chinese) input.
Auto-complete	Dynamically suggests relevant tags as you type.
Negative Prompt	Defines what you don’t want in the image. These carry higher priority than positive prompts.
Quality Tags	A preset library for frequently used enhancement tags.

Prompt Tools Guide : Supercharge Your AI Art: 3 Must-Try Prompt Tools on PixAI

Generation Parameters

Parameter	Description
Steps	Determines the level of detail. Higher steps yield more refinement but cost more credits.
CFG Scale	Controls prompt adherence. Higher values follow instructions strictly; values that are too high cause artifacts. Recommended: 5–7.
Seed	Determines the initial noise. Using the same seed with similar prompts results in nearly identical images—ideal for fine-tuning.
Ratio	The aspect ratio of the output. Choose based on your desired composition.
Quantity	Number of images generated per batch. Useful for picking the best result from multiple variations.

Advanced Features

Feature	Description	Notes
Face Fix	Fixes distorted facial features.	May cause artifacts on older models.
Enhance Details(HiRes)	Upscales resolution and adds detail.	See HiRes Upscaling Guide.
Reference Image	Guides the composition.	Lower values = closer to original. (0 = exact, 1 = no reference).
Character Ref	Maintains character consistency.	Effectiveness varies by model.
ControlNet	Precisely controls composition (poses, line art, etc.).	Advanced users only.

Panel Introduction: Your Creative Command Center: A Guide to PixAI’s Generation Panel

Key Concepts in Prompt Engineering

Use Tags from Established Databases

Most popular SDXL models on PixAI are trained on Danbooru or e621 datasets. For the most stable results, use tags found in these databases.

💡 While CLIP (the model’s text encoder) understands some natural English (e.g., “red chess piece”), formal tags are significantly more reliable.

Chinese Resources

[Stable Diffusion雜談] Stable Diffusion 與 CLIP 的基本原理 — 工gin師
一次掌握CLIP：AI跨模態理解的關鍵技術 — Josh的沙龍
[Day29] SD XL：SD的限界突破 — Nick・iT邦鐵人賽

English Resources

CLIP: Connecting text and images — OpenAI Official Blog
OpenAI CLIP Model Explained — Lightly
Two Text Encoders in SDXL 1.0 — myByways
SDXL Official Documentation — Hugging Face

Common Errors:

❌ narrow eyes → Incorrect (Use narrowed eyes).
❌ young man → Non-existent (Use 1boy + specific traits).
❌ handsome → Subjective/Non-existent (Use bishounen or manly).

Spaces Over Underscores

While databases like Danbooru use underscores (black_hair), PixAI processes prompts better with spaces (black hair).

Tool Spotlight: PixAI Tagger

Unsure how to describe an image? Use the PixAI Tagger to reverse-engineer any visual:

PixAI Tagger Demo ➚

Upload an image, and the AI will list the exact tags needed to recreate that style or subject.

Prompt Structure & Weighting

The sequence of your tags matters. Follow this hierarchy:

[Core Subject], [Details], [Camera/Framing], [Art Style], [Quality Tags]

Example:

1boy, black hair, short hair, blue eyes,
upper body,
anime style,
masterpiece, best quality

Order	Category	Examples
1	Core Subject	`1boy`, `1girl`, `scenery`
2	Subject Details	`black hair`, `blue eyes`, `smile`
3	Camera / Framing	`upper body`, `close-up`, `from below`
4	Art Style	`anime style`, `illustration`, `realistic`
5	Quality Tags	`masterpiece`, `best quality`

💡 Quick Tip: Separate each element with a comma. While you can use line breaks to organize your prompts, they will be ignored during the actual generation process.

Using Weights

Control how much influence a specific tag has on the result.

Use (keyword:weight) from 0.1 to 2.0.

< 1.0: Decreases emphasis.
> 1.0: Increases emphasis.

(black hair:1.2)    # Slight boost
(blue eyes:1.4)     # Strong boost
(smile:0.8)         # Slight reduction

⚠️ Note: Avoid going over 1.5 to prevent “deep-frying” the image.

Bracket Weighting (Advanced)

Syntax	Example	Effect
`(keyword)`	`((black hair))`	Each layer ×1.1; double layer = 1.21
`{keyword}`	`{{black hair}}`	Each layer ×1.05; double layer = 1.1025
`[keyword]`	`[[black hair]]`	Each layer ÷1.05; double layer ≈ 0.907

💡 Pro Tip: Using the (keyword:weight) format is generally more intuitive and easier to control.

Escape Characters

If your prompt requires literal parentheses (for example, in a character’s name), use a backslash to “escape” them:

\(special_character\)

However, this is rarely necessary in most common use cases.

Arithmetic Operators (⚠️ Not Applicable to PixAI)

⚠️ Important Reminder: The following symbols have been verified by the community to be ineffective on PixAI. These are features specific to local tools like A1111 (Automatic1111 WebUI) and are not supported by PixAI.

📖 To learn more about which syntaxes work or do not work on PixAI, you can refer to the experimental article “Using Emphasis on PixAI” compiled by Maki in the PixAI Official Discord.

Operator	Example	Potential Effect	Status on PixAI
AND	`a cat AND a dog`	Satisfies two conditions simultaneously	❌ A1111 feature; ineffective on PixAI
+	`a cat + a dog`	Satisfies two conditions simultaneously	❌ No confirmed effect
\|	`red\|blue hair`	Randomly selects one of the options	❌ ComfyUI OR syntax; ineffective on PixAI
_	`coffee_cake`	Connects words into a single concept	⚠️ Usually equivalent to a space, unless the tag itself contains an underscore

💡 Conclusion: On PixAI, please use the standard method of separating tags with commas (,) directly, and avoid using these operators.

Art Style Tag Reference

Common style tags you can choose based on your desired look:

English Tag	Effect
`illustration`	Polished and refined illustration style
`painting`	Traditional painted texture and brushwork
`anime`	Classic Japanese animation style
`comic`	Western comic book or graphic novel style
`digital art`	Modern digital painting and clean lines
`game CG`	High-end cinematic or game cinematic style
`semi-realistic`	A hybrid blend of anime features and realistic textures
`2.5D`	2D stylized art with a sense of 3D depth and volume
`realistic`	An appearance that closely mimics real life
`photograph`	High-fidelity, camera-realistic photo effect

Quality Tag Reference

Universal Quality Tags

masterpiece, best quality, ultra-detailed, 8k

Recommended Tags by Style

Style	Tag
Anime	`masterpiece, best quality`
Intricate Details	`ultra-detailed, extremely detailed`
Realistic Style	`photorealistic, extremely detailed, 8k`
3D Rendering	`unreal engine rendered, 3D rendered`

💡 Quick Tip: More is not always better. Usually, 2–3 quality tags are sufficient. Overloading them can lead to over-sharpening or distorted details.

Negative Prompt Essentials

Negative prompts tell the AI what not to include in the image.

Standard Universal Negatives

lowres, bad anatomy, bad hands, text, error,
missing fingers, extra digit, fewer digits,
cropped, worst quality, low quality,
normal quality, jpeg artifacts, signature,
watermark, username, blurry

Essential for Male Characters

1girl, female, breasts

Essential for Landscapes

1girl, 1boy, person, human, figure, character

For more in-depth information, please refer to the dedicated chapters for each specific topic.

Quick Start Examples

Example 1: Male Character

Positive:

1boy, solo, male focus,
short hair, black hair, blue eyes, smile,
upper body, looking at viewer,
simple background,
masterpiece, best quality

Negative:

1girl, female, breasts, lowres, bad anatomy

Example 2: Landscape