SDXL Prompt Guide: Formula, Examples & Settings (2026)

SDXL reads prompts differently from SD 1.5. It understands short descriptive sentences, needs fewer "masterpiece, best quality" tags, and is far more sensitive to resolution. This guide gives you a formula, ready prompts and the settings that matter.

The SDXL prompt formula

[subject], [action or pose], [setting], [lighting], [camera or medium], [style details]

Put the most important thing first and describe it concretely. "An old fisherman mending a net on a wooden pier" works better than "fisherman, old, net, pier, detailed, 8k".

A full example:

an old fisherman mending a red net on a wooden pier, early morning fog, soft diffused light, 85mm lens, shallow depth of field, muted film colors

Copy-paste SDXL prompts

Each prompt is paired with a short negative prompt. For longer lists, see the negative prompt guide.

Portrait photo

close-up portrait of a woman in her 40s with freckles and grey-streaked hair, window light from the left, neutral background, 85mm lens, natural skin texture, documentary photography

cartoon, illustration, 3d render, plastic skin, airbrushed, oversaturated, watermark, text

Product shot

a matte black ceramic coffee mug on a light oak table, soft morning light, clean beige background, product photography, centered composition, subtle shadow

cluttered background, dust, fingerprints, harsh reflections, text, watermark, lowres

Landscape

alpine lake at sunrise, mist over the water, pine forest on the shore, snow-capped peaks reflected in the lake, wide-angle landscape photography, crisp detail

people, buildings, power lines, oversaturated, blurry, text, watermark

Interior

a small Scandinavian living room, white walls, light wood floor, linen sofa, large window with afternoon sun, potted fig tree, architectural photography

distorted perspective, warped lines, clutter, people, text, watermark

Illustration

a fox reading a book under a mushroom in the rain, children's book illustration, watercolor and ink, soft pastel palette, white paper texture

photograph, photorealistic, 3d render, dark, gritty, text, watermark

Anime style

a girl with short silver hair standing on a rooftop at dusk, city lights below, wind in her coat, anime style, cel shading, vivid sunset colors

photorealistic, 3d, lowres, bad anatomy, extra fingers, text, watermark

Fantasy scene

an ancient stone bridge over a misty gorge, a lone traveler with a lantern, towering cliffs, epic fantasy concept art, volumetric light, muted greens and greys

modern buildings, cars, cartoon, oversaturated, text, watermark

Base SDXL produces these styles on its own. Fine-tuned checkpoints (photo, anime, illustration) usually push one style much further; check each model page for its preferred prompt format.

Settings that matter more than words

Resolution

SDXL works best at about one megapixel. The Diffusers documentation says most SDXL checkpoints work best at 1024×1024; 768 and 512 work but look worse, and anything below 512 is not recommended.

These sizes keep the pixel count close to 1024 × 1024. Generating at 512 × 512, a habit from SD 1.5, is the most common reason SDXL images look soft or broken.

Guidance (CFG) and steps

Diffusers uses a default guidance scale of 5.0 for SDXL. A range of about 4–8 is a reasonable starting point; very high values tend to burn colors and contrast. The Diffusers docs also note that DPM++ samplers below 50 steps can produce visual artifacts with SDXL and suggest enabling Karras sigmas, which most UIs list as "DPM++ 2M Karras".

Long prompts

The CLIP text encoders read 75 tokens at a time. AUTOMATIC1111 and similar UIs split longer prompts into chunks automatically, but words near the start of the prompt still tend to matter most. Keep the subject first.

What is different from SD 1.5

• Two text encoders. SDXL uses CLIP ViT-L and OpenCLIP ViT-bigG. In Diffusers you can send a second prompt to the second encoder with prompt_2 (for example, the subject in prompt and the style in prompt_2). If you leave it empty, both encoders get the same prompt, which is what most UIs do.
• Fewer quality tags. Lists like "masterpiece, best quality, ultra detailed, 8k" were a workaround for SD 1.5. In SDXL they add little; spend those words on the subject, light and medium.
• Sentences work. Short natural phrases ("a dog sleeping on a red sofa") are understood. Tag lists still work, especially on anime checkpoints trained on tags.

The refiner: optional

SDXL 1.0 shipped with a separate refiner model that adds fine detail in the last, low-noise part of generation. In the Diffusers "ensemble of experts" setup, the base model runs the first 80% of denoising (denoising_end=0.8) and the refiner runs the rest (denoising_start=0.8).

The refiner needs extra memory and time. Many fine-tuned SDXL checkpoints are published to be used without it, so check the model page before adding it.

Making a word stronger

In AUTOMATIC1111 and ComfyUI, (word:1.2) increases a word's weight and (word:0.8) reduces it:

a cottage in a meadow, (wildflowers:1.3), (fog:0.7), golden hour

Stay between about 0.7 and 1.4. Larger values often distort the image. This syntax does not work in Midjourney or in plain API calls.

What SDXL still struggles with

The official SDXL model card lists these limitations:

• It cannot render legible text. Add text afterwards in an editor.
• Faces and people may not be generated properly, especially small faces in wide shots.
• Complex compositions ("a red cube on top of a blue sphere left of a green cone") are hard.
• It does not reach perfect photorealism.

For layout problems, regional prompting gives each part of the image its own prompt. For variety across many images, wildcards swap words automatically. Useful add-ons are covered in the extensions guide.

Build your prompt

The prompt generator assembles subject, style, lighting and camera terms into one prompt with a matching negative field. It runs in your browser and does not generate images.

Primary sources

• Stability AI: SDXL base 1.0 model card
• Hugging Face Diffusers: Stable Diffusion XL
• SDXL paper (Podell et al., 2023)
• AUTOMATIC1111 Web UI features
• ComfyUI README (prompt weighting)