Stable Diffusion & SDXL Prompts - Weights & LoRA

Essential Stable Diffusion / SDXL prompt templates — six-segment structure, weight syntax, LoRA, sampling params. Copy a template, fill the variables, and ship.

AI Prompts·30 commands·Last updated 2026-08-21

Six-Segment Structure 5

[Subject] + [Action/Pose]

Subject + pose first, highest weight

[Environment]

Environment/scene description

[Lighting/Mood]

Lighting and mood

[Camera/Technical]

Camera/technical (focal length, DoF)

[Artist Style]

Artist style at the end

Weight Syntax 5

(keyword:1.3)

Boost a token (>1 raises weight)

(keyword:0.7)

Reduce a token (<1 lowers weight)

((word))

~1.1x mild boost, nestable

Balance weights; avoid one token too high

Too high → oversaturation/artifacts

Use weights to spotlight subject, weaken bg noise

Focus the composition

Negative Prompts 5

blurry, low quality, watermark

General negatives: blur/low-q/watermark

distorted hands, extra fingers, bad anatomy

Common human defects

oversaturated, jpeg artifacts

Oversaturation/compression artifacts

duplicated, mutated, deformed

Structural anomalies

Add style-specific negatives (e.g. drop conflicting words)

Style consistency

LoRA & ControlNet 5

<lora:name:0.6>

Mount a LoRA; 0.6-0.9 is the sweet spot

Stack at most 3 LoRAs to avoid clashes

Too many fight each other

ControlNet: OpenPose / Canny / Depth

Pose/lineart/depth control

img2img denoising 0.3-1.0

Lower = closer to the source

Tune LoRA weight per base model

Sweet spot differs by base

Sampling Params 5

CFG Scale 6-8

Prompt adherence; too high → artifacts

Sampling Steps 25-30

Enough denoising steps

Sampler: DPM++ 2M Karras

Common high-quality SDXL sampler

Native 1024×1024

SDXL optimal training resolution

Minimum 8GB VRAM

VRAM floor reference

Templates 5

portrait of {subject}, studio lighting, 85mm, sharp focus, <lora:film:0.7>

Portrait template

{scene}, cinematic lighting, wide shot, detailed, art by {artist}

Scene illustration template

product photo of {item}, softbox, white background, 4k

Product shot template

negative: blurry, low quality, watermark, extra fingers

General negative template

((masterpiece)), {subject}, intricate details, volumetric light

High-quality boost template

Tips

SDXL prompts want "six segments + balanced weights": subject first, style last; weights >1.4 cause oversaturation/artifacts, LoRA 0.6-0.9 is safest.
Negative prompts are a quality floor: blurry/low quality/watermark plus distorted hands/extra fingers for people are nearly mandatory.
CFG 6-8, Steps 25-30, DPM++ 2M Karras is the safe SDXL default; native 1024×1024 is most stable — avoid non-native resolutions.
For precise pose/composition use ControlNet (OpenPose/Canny/Depth) — more controllable than words alone.

FAQ

What happens if CFG Scale is too high?

You get artifacts: oversaturation, harsh contrast and broken structure. The usual range is 6 to 8, enough to follow the prompt without hurting image quality. Below 5 the prompt is diluted and the model wanders; above 10 visible artifacts are almost guaranteed. When tuning, fix the seed and step count and change only CFG, otherwise you cannot attribute the difference.

Why list style-conflicting terms in the negative prompt?

The model treats the prompt as a set of tendencies, so if photorealistic appears on the positive side you should exclude conflicting styles such as cartoon from the negative side. Otherwise the two pull against each other and you get a muddy result that is neither realistic nor cartoonish. For people, keep distorted hands, extra fingers and bad anatomy in the negative prompt to suppress structural defects. More is not better — negatives should match the style.

How many LoRAs can I stack?

Keep it to three at most; beyond that the styles fight each other and the output averages out rather than combining. Weights around 0.6 to 0.9 are the common sweet spot, but the right value depends on the base model — the same LoRA needs retuning on a different base. Retest the weight when you switch bases instead of carrying the old value over.

How many sampling steps should I use?

25 to 30 is the usual range and enough for the model to finish denoising. Going higher adds little while increasing time linearly; below 20 visible detail is lost and the image looks soft. When tuning, hold everything else fixed — seed, CFG, LoRA weights — so you can attribute the difference.

How should I think about the img2img denoising value?

It expresses how much of the original is retained, from 0.3 to 1.0. Near 0.3 the result hugs the source with only light adjustment, suited to retouching and local fixes. Near 1.0 the source barely matters and it is effectively a fresh generation. The middle range of 0.5 to 0.7 keeps the composition while changing style.

Official References

Each command links to its official documentation below, so you can verify the latest usage and read deeper.

Maintained by LaoHand

Publicly updated on Aug 21, 2026, continuously proofread against official docs.

Contact Us

Wrong command or description? Send us corrections, business inquiries or product feedback by email.