Skip to content
Log inSign up free

Prompting guide

How to prompt Stable Diffusion XL

Stable Diffusion XL prompt guide: how to write short SDXL prompts, use weights and negative prompts, get past the 75-token limit, with SDXL prompt examples.

Try Stable Diffusion XLStable Diffusion XL model page

Mountain sunrise
Made with Stable Diffusion XLMountain sunrise

Quick start

Your first great Stable Diffusion XL prompt

Everything a beginner needs in one minute: what to know first, the rules that matter most, a template to fill in, and prompts to copy.

What to know first

Stable Diffusion XL wants short phrases separated by commas, with the subject first, not chatty sentences. Mage's default SDXL checkpoint is tuned for photos, so photographic words (a lens, the light, "photo") pull hardest. At the default Prompt Behavior it reads only about the first 75 tokens, 50 to 60 words, and drops the rest. The Negative Prompt and weights both work.

The rules that matter most

  1. Subject first, in short phrases

    The subject and count, then appearance, pose, framing, setting, and style, in that order, separated by commas.

  2. Stay under about 60 words

    Past that, the default Prompt Behavior drops the rest. For a longer prompt, set Prompt Behavior to V2 (Unlimited).

  3. Speak like a photographer

    A lens (85mm, wide angle), a light (soft backlight, golden hour), and the word photo steer the default checkpoint best.

  4. Weight one detail, gently

    (neon reflections:1.2) stresses one phrase. Mage caps weights at 1.6; stay between 1.1 and 1.4.

  5. Use the Negative Prompt

    List what you never want to see: blurry, cartoon, text, watermark, deformed hands.

Fill in this template

Replace each part in square brackets with your own idea, and leave out a part you do not need.

Template
[subject], [key appearance], [pose or action], [framing and lens], [setting], [light], [style]
The template, filled in
photo of a young woman, yellow raincoat, wet hair, standing, looking over her shoulder, medium shot, 50mm lens, shallow depth of field, wet city street at night, (neon reflections:1.2), cinematic color, film grain
  • [subject] Who or what, with the count: photo of a young woman, two horses.
  • [key appearance] Two or three visible details: yellow raincoat, wet hair.
  • [pose or action] standing, looking over her shoulder.
  • [framing and lens] medium shot, 50mm lens, shallow depth of field.
  • [setting] wet city street at night.
  • [light] The light, weighted if it matters most: (neon reflections:1.2).
  • [style] cinematic color, film grain.

Prompts to start from

  • A food close-up

    close-up photo of a stack of blueberry pancakes, maple syrup dripping down the side, butter melting on top, rustic wooden table, morning window light, shallow depth of field, 100mm macro

    Why it works: The subject leads, every phrase is something visible, and the lens and light do the styling. It fits well inside the 75-token window.

  • A wide interior

    wide angle photo of a minimalist concrete living room, low linen sofa, floor-to-ceiling window, pine forest outside, overcast daylight, architectural photography, neutral tones

    Why it works: "wide angle" and "architectural photography" tell the photoreal checkpoint exactly what kind of shot to take.

  • A street portrait with one weight

    portrait photo of an old street musician playing violin, gray beard, flat cap, eyes closed, city square at dusk, (warm streetlight:1.2), 85mm, bokeh

    Why it works: One weighted light carries the mood. Pair it with the Negative Prompt cartoon, illustration, blurry, deformed hands.

  • Wildlife in motion

    photo of a snowy owl in flight over a frozen field, wings fully spread, low winter sun, telephoto lens, fast shutter, crisp feathers

    Why it works: Telephoto and fast shutter are the words a wildlife photographer would use, and the checkpoint answers to them.

Advanced

Advanced Stable Diffusion XL prompting

Techniques for more control once the basics work.

Token order and the 75-token window

SDXL reads a prompt in tokens, small pieces of words. At the default Prompt Behavior, V1 (Legacy), it keeps only about the first 75: 50 to 60 words. Anything after that is silently dropped, so a perfect detail at the end of a long prompt never arrives.

Put the subject and its most important details first, the setting and style after. When you truly need more words, set Prompt Behavior to V2 (Unlimited), which reads long prompts in chunks.

  • Short tag clusters beat sentences: "red scarf, snowy street" says more per token than "she is wearing a red scarf while walking down a street covered in snow".
  • Skip filler such as masterpiece, best quality, or stunning. It spends tokens without adding anything you can see.
  • Say the count plainly: one woman, two dogs, a crowd.

Weights

Weights work in the prompt and in the Negative Prompt. Use them sparingly: one or two weighted phrases between 1.1 and 1.4 nudge the image, while many weights fight each other and distort it.

WriteWhat it does
(phrase:1.3)Stresses the phrase by that factor. Mage caps weights at 1.6.
(phrase)A little more weight, about 1.1.
[phrase]A little less weight.
\( \)Literal parentheses, read as text instead of a weight.

The negative prompt

The Negative Prompt is where most of the clean-up happens. List what keeps appearing that you do not want, as short phrases: blurry, lowres, text, watermark, cartoon, illustration, deformed hands, extra fingers.

Put the opposite of your style there. For a photograph, cartoon, illustration, painting; for an illustration on an art checkpoint, photo, realistic.

  • Keep it to what you actually see going wrong. A giant stock list dilutes the parts that matter.

Checkpoints, LoRAs, and anime

Mage's default SDXL checkpoint is a photoreal fine-tune, which is why photographic words work so well. The Model setting can switch to another SDXL checkpoint, and a checkpoint trained on anime or art answers to its own vocabulary.

For an anime checkpoint, lean harder into tags: a count tag first (1girl, 1boy, 2girls, solo), then hair, eyes, outfit, expression, pose, framing, and background. For anime without choosing a checkpoint, Anima is built for it.

  • A LoRA added in the settings often answers to a trigger word named by its maker. Put that word near the start of the prompt.

A tag prompt for an anime checkpoint

1girl, solo, anime illustration, long black hair, blue eyes, school uniform, soft smile, standing under cherry blossoms, spring petals, full body, looking at viewer, clean lineart, detailed background

Why it works: Count tags first, then concrete tags in a fixed order: the vocabulary anime checkpoints learned. On the default photoreal checkpoint, the same tags come out half real.

Starting from an image

Upload a picture in the Image setting and SDXL starts from it: the composition and colors carry over and the details change. Prompt Strength sets the balance; higher listens to the prompt more and the picture less.

Describe the whole finished picture. SDXL does not follow edit instructions such as "change the sofa to leather".

  • ControlNet keeps one structure from a picture, its pose, edges, or depth, while the prompt changes everything else.
  • Style References borrow a look, and Face References a face, from images you upload. The weight beside each sets how strongly.

Common mistakes

Avoid this, do this instead

  • Avoid: A 100-word prompt at the default Prompt Behavior.

    Instead: Under about 60 words, or Prompt Behavior set to V2.

  • Avoid: A chatty request.

    Could you make a picture of a cute little dog sitting in the grass in a park on a really sunny day?

    Instead: Comma-separated visual phrases.

    photo of a corgi puppy sitting in green grass, city park, sunny afternoon, low angle, 35mm, shallow depth of field

  • Avoid: Precise arrangements, such as "a red cube on top of a blue sphere".

    Instead: A simpler composition, or Stable Diffusion 3.5 Large or HiDream for exact placement.

  • Avoid: Heavy weights such as (eyes:2.5).

    Instead: One or two weights between 1.1 and 1.4; Mage caps them at 1.6 anyway.

  • Avoid: Edit instructions with an uploaded picture.

    Instead: A description of the whole result, with Prompt Strength set to taste.

  • Avoid: Words on a sign.

    Instead: No text in the prompt; add the lettering afterwards.

FAQ

Stable Diffusion XL prompting questions

How long can an SDXL prompt be?
About 75 tokens, or 50 to 60 words, at the default Prompt Behavior; the rest is dropped. Set Prompt Behavior to V2 (Unlimited) for longer prompts.
What should I put in an SDXL negative prompt?
What keeps going wrong, as short phrases: blurry, lowres, text, watermark, deformed hands, extra fingers. Add the styles you do not want, such as cartoon and illustration for a photo.
How do prompt weights work in SDXL?
(phrase:1.3) multiplies its weight by 1.3, (phrase) adds a little, and [phrase] takes a little away. Mage caps weights at 1.6. Weight one or two phrases at most.
Can SDXL write text?
No. SDXL cannot render legible text. Add lettering afterwards, or use a model that writes text, such as GPT Image or Mango.
How do I make anime with SDXL?
Pick an anime SDXL checkpoint in the Model setting and write tags: a count tag such as 1girl, then hair, eyes, outfit, pose, framing, and background. Or use Anima, which is built for anime.
Should I use Stable Diffusion XL or SDXL Plus?
They share the photoreal checkpoint and the prompt style. SDXL Plus adds a high-resolution pass and automatic face and hand repair, so it is slower and better for people. Plain SDXL is quicker for landscapes, animals, and drafts.

Make your first image with Stable Diffusion XL

New accounts get 300 free Gems, enough to try Stable Diffusion XL. No install, no credit card.

Try Stable Diffusion XL