Skip to content
Log inSign up free

Prompting guide

How to prompt Stable Diffusion v1.5

Stable Diffusion 1.5 prompt guide: how to write short SD 1.5 prompts, negative prompts, and weights, match older work, with SD 1.5 prompt examples.

Try Stable Diffusion v1.5Stable Diffusion v1.5 model page

Long exposure seascape
Made with Stable Diffusion v1.5Long exposure seascape

Quick start

Your first great Stable Diffusion v1.5 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 1.5 is the classic open model: cheap, fast, and still the way to match work made on it. It learned from small images with mostly English captions, so keep prompts short, literal, and in English: a subject, a setting, a light, and a medium. The Negative Prompt does much of the clean-up. It cannot write legible text, and it struggles with complex arrangements.

The rules that matter most

  1. Keep it short and literal

    One subject, one setting, one light: "a photo of an old fisherman, harbor, morning light" beats a paragraph.

  2. Write in English

    It learned mostly from English captions, so prompts in other languages come out much weaker.

  3. Name the medium

    a photo of, an oil painting of, a watercolor of: at the start, or "photo" at the end.

  4. Lean on the Negative Prompt

    blurry, lowres, bad anatomy, extra fingers, text, watermark: what you never want to see.

  5. Keep the composition simple

    One main subject, placed plainly. Stacked or intertwined objects confuse it.

Fill in this template

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

Template
a [medium] of [subject], [one detail], [setting], [light], [framing]
The template, filled in
a photo of a red vintage bicycle, wicker basket full of sunflowers, leaning against a stone wall, golden hour, side view
  • [medium] photo, oil painting, watercolor painting, pencil sketch.
  • [subject] One main subject: a red vintage bicycle.
  • [one detail] The detail that matters most: wicker basket full of sunflowers.
  • [setting] leaning against a stone wall.
  • [light] golden hour, overcast, candlelight.
  • [framing] side view, close-up, wide shot.

Prompts to start from

  • The classic demo, extended

    a photo of an astronaut riding a horse on mars, red desert, dramatic sky, wide shot

    Why it works: The model card's own example, with three short phrases added: the subject first, then the place, the sky, and the framing.

  • An oil painting

    an oil painting of a red fox asleep in tall grass, golden hour, impressionist brushstrokes, warm palette

    Why it works: The medium opens the prompt, and the brushstroke and palette words reinforce it.

  • A simple still life

    a photo of a vintage typewriter on a wooden desk, a cup of black coffee beside it, morning window light, shallow depth of field

    Why it works: Two objects side by side is an arrangement 1.5 handles well; the light and the depth of field do the styling.

  • A watercolor landscape

    a watercolor painting of a sailboat on a calm lake at sunrise, soft pastel colors, mountains in the distance

    Why it works: One subject, one setting, one palette: literal, short, and easy for 1.5 to get right.

Advanced

Advanced Stable Diffusion v1.5 prompting

Techniques for more control once the basics work.

Short, literal, front-loaded

1.5 reads a prompt in chunks of about 75 tokens, 50 to 60 words, and Mage passes a longer prompt through in several chunks. It still responds best to a compact one: the subject and its key detail first, then the setting, light, and medium, as short phrases separated by commas.

Describe what can be seen, in plain words. Story, abstract ideas, and mood words with no visual behind them give it little to draw.

  • Count small: one or two subjects. Crowds and exact numbers drift.
  • The default checkpoint needs no quality words. Some community checkpoints were trained to expect tags such as best quality; add them only on those.

Negative prompt, weights, and Clip Skip

The Negative Prompt matters more on 1.5 than on newer models. A typical one: blurry, lowres, bad anatomy, extra fingers, deformed, text, watermark. Add the opposite of your style: cartoon, illustration for a photo; photo, realistic for a painting.

Weights work in both prompts: (phrase:1.3) stresses a phrase, (phrase) adds a little, [phrase] takes a little away. Mage caps weights at 1.6. Write literal parentheses as \( and \).

  • Clip Skip is on by default, which most community checkpoints, anime ones especially, expect. If an image from an older workflow does not match, try turning it off.

Checkpoints and matching older work

The Model setting picks the 1.5 checkpoint, and Mage's default is a general fine-tune. Community checkpoints each have their own taste: anime ones want tags (1girl, solo, long hair), photoreal ones want camera words.

To match an image made on 1.5 elsewhere, choose the same checkpoint and LoRAs, paste the same prompt and Negative Prompt, and match the Steps, Guidance Scale, Scheduler, and Seed. Expect a close match rather than an identical one, because every platform differs in small ways.

  • A LoRA usually answers to a trigger word named by its maker. Put it near the start of the prompt.

A tag prompt for an anime checkpoint

1boy, solo, short silver hair, red eyes, black hooded cloak, holding a lantern, snowy forest at night, upper body, looking at viewer

Why it works: The count tag first, then hair, eyes, outfit, prop, place, and framing as tags: the vocabulary anime 1.5 checkpoints were trained on.

Images, poses, and faces

Upload a picture in the Image setting to start from its composition and colors, with Prompt Strength setting the balance. The prompt describes the whole result, not an edit.

ControlNet holds one structure from a picture, its pose, edges, or depth, while the prompt changes everything else. Style References and Face References borrow a look or a face from images you upload.

  • Keep Hires on: the second, larger pass helps faces and fine detail.

What to take to another model

1.5 cannot render legible text, struggles with arrangements such as "a red cube on top of a blue sphere", and can malform faces and hands, small ones especially. For more detail, move up to Stable Diffusion XL or Z-Image; for exact placement, Stable Diffusion 3.5 Large or HiDream; for text, GPT Image or Mango.

Common mistakes

Avoid this, do this instead

  • Avoid: A long, poetic sentence.

    A melancholy old lighthouse keeper who has spent his whole life alone watches the storm he always feared

    Instead: Short phrases for what can be seen.

    a photo of an old lighthouse keeper at a window, gray beard, storm outside, rain on the glass, lamp light

  • Avoid: A prompt in another language.

    Instead: The same prompt in English.

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

    Instead: One main subject placed simply, or Stable Diffusion 3.5 Large for arrangements.

  • Avoid: Words on a sign.

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

  • Avoid: An empty Negative Prompt.

    Instead: A short list of what you never want to see.

  • Avoid: Edit instructions with an uploaded picture.

    Instead: A description of the whole result.

FAQ

Stable Diffusion v1.5 prompting questions

Is Stable Diffusion 1.5 still worth using?
For matching older work, for cheap images at volume, and for its huge range of community checkpoints, yes. For detail, text, or precise layouts, newer models do better.
How long should a Stable Diffusion 1.5 prompt be?
Short: about 10 to 30 words. Mage reads longer prompts in chunks, but 1.5 follows a compact, literal prompt best.
What negative prompt should I use with SD 1.5?
Start with blurry, lowres, bad anatomy, extra fingers, deformed, text, watermark, then add the styles you do not want, such as cartoon for a photo.
Why do faces look distorted in SD 1.5?
Small faces have few pixels to work with. Frame closer, keep Hires on, and add bad anatomy and deformed face to the Negative Prompt. For people at a distance, SDXL Plus repairs every face automatically.
Can Stable Diffusion 1.5 write text?
No, it cannot render legible text. Add lettering afterwards, or use GPT Image or Mango.
Can I recreate an image I made with SD 1.5 somewhere else?
Closely, yes. Choose the same checkpoint and LoRAs, use the same prompt and Negative Prompt, and match the Steps, Guidance Scale, Scheduler, and Seed. Small differences between platforms mean the result is close rather than identical.

Make your first image with Stable Diffusion v1.5

New accounts get 300 free Gems, enough to try Stable Diffusion v1.5. No install, no credit card.

Try Stable Diffusion v1.5