Effective Stable Diffusion Prompts: Method, Weights, and Settings

A prompt copied from a forum produces a stunning image for its author and a muddy mess for you. The cause is rarely the text. Stable Diffusion isn't a single tool, it's a family of models, used through several interfaces, with settings you control yourself. The same prompt doesn't have the same effect on a base model and on a community model trained for anime.

That's also what makes it interesting. Where Midjourney decides a lot of things for you, Stable Diffusion leaves you in charge of almost everything, and a well-documented image can be reproduced almost identically. You still have to write for the right model.

Identify your model before writing a word

How you write depends first on the loaded model, often called a checkpoint. The advice you read online contradicts itself because it's not talking about the same models.

For a community model, the download page generally lists recommended settings, trigger words, and sometimes example prompts. Read it: it's worth more than any general guide, this one included.

Generating far from the base resolution is a classic mistake. An SD 1.5 model run directly at a very large size tends to repeat subjects, and an SDXL model pushed too low loses quality. Stability AI's documentation also notes that going outside the recommended sizes for SDXL can cause unexpected framing. For a larger image, generate at the native size, then upscale.

A note on FLUX, which you'll run into in the same interfaces: it's not a Stable Diffusion model, and Black Forest Labs states that it doesn't support negative prompts. A good part of what follows therefore doesn't apply to it directly.

Write the positive prompt like a shot sheet

Take an accountant who wants to illustrate an article about the end of the fiscal year.

Weak version: beautiful office, professional, masterpiece, best quality, ultra detailed

The model mostly receives noise. It has no idea what object to show, and quality words, useful on some community models, don't make up for a missing subject.

Corrected version:

photograph of a tidy wooden desk with stacked paper invoices, a fountain pen and a mechanical calculator, late afternoon sunlight through a window, long soft shadows, shallow depth of field, muted beige and dark green tones, 50mm lens

This time, the subject comes first, followed by the setting, the light, the palette, and the lens. Put what matters most at the beginning: interfaces split long prompts, and an essential element lost at the end of the text is less likely to be respected.

On AUTOMATIC1111 and its derivatives, the prompt is processed in blocks of 75 tokens. You can go past that limit, with the text then spread over several blocks, but the split can land in the middle of an idea. The keyword BREAK, written in capitals, forces a new block at the spot you choose. It's mostly useful for separating two ideas, like the subject and the background.

Write in English: it's the language these models respond to best. If English isn't your first language and it slows you down, draft in your own language and have it translated, then proofread it.

Weights: the exact syntax

A word's weight tells the model how much attention to give it. The following syntax comes from the AUTOMATIC1111 documentation, and ComfyUI shares the parenthesis form.

Stay between 0.8 and 1.3 to begin with. Beyond that, the image tends to degrade, and the overloaded term takes over the composition. If you have to push a word to 1.6 to make it show up, the problem probably lies elsewhere: word order, a poorly suited model, or a contradiction in the prompt.

Watch out for square brackets. In this guide's prompts, [VARIABLE] indicates a value to replace. Left as-is in AUTOMATIC1111, they'd be interpreted as a drop in attention.

AUTOMATIC1111 also offers a notation for switching prompts mid-generation. [fantasy:cyberpunk:16] draws a fantasy scene, then switches to cyberpunk from step 16, and [cow|horse] alternates the two words at each step. These two syntaxes are handy for blending concepts, but they don't work as-is in ComfyUI without an extension, and the documentation specifies that they don't apply to additional networks like LoRAs.

The settings that matter, and the ones that can wait

The same text changes radically from one combination of settings to another. The values below are starting points from common practice, not official rules. Your model's page always takes precedence.

CFG scale. It sets how closely the image follows the prompt. Too low, and the model drifts away from your request. Too high, and colors oversaturate and edges harden. The ComfyUI documentation sums up this tradeoff well: raising this setting brings the image closer to the prompt, but an excessive value hurts quality. The community often starts at 5 to 7 on SDXL, more like 7 to 9 on SD 1.5, and lower on SD 3.5.

Steps. The number of denoising passes. Past a certain threshold, you pay in compute time without real gain: compare 20, 30, and 40 on your model, then keep the lowest that satisfies you. The official examples for SD 3.5 run between 28 and 40 steps.

Sampler. It determines the denoising method. Euler a and DPM++ 2M Karras remain common choices. Don't spend your evening on this at first: the sampler changes the result less than the prompt or the model.

Seed. The starting point of the noise. By fixing it, you can change a single word and see its real effect, which is the basis of any serious comparison.

Negative prompt. It steers the model away from what you list. It's useful in moderation, and the long lists copied from old tutorials are rarely the right solution. The article on the negative prompt explains how to dose it and when it stops working. On ComfyUI, it goes through a second text-encoding node, connected to the KSampler's negative input.

Three complete prompts to adapt

Replace each variable with English text, remove the brackets, paste each part into the matching field, and start with the settings recommended for your model.

For a dish photo for a restaurant owner, on SDXL:

Positive:
overhead photograph of [DISH] served on [PLATE TYPE], on a [MATERIAL] table, soft natural side light from a window, visible steam, realistic food texture, [2 COLORS] tones, shallow depth of field, editorial food photography

Negative:
text, watermark, logo, blurry, plastic-looking food

Settings: [WIDTH] x [HEIGHT] at a native size for the model, CFG 6, 30 steps, DPM++ 2M Karras, fixed seed

For a textbook or activity-sheet illustration, useful to a teacher, on SDXL:

Positive:
educational illustration of [CONCEPT OR OBJECT TO EXPLAIN], clean line art with flat colors, simple shapes, palette of [3 COLORS], white background, single centered subject, children's science book style

Negative:
photo, text, letters, realistic shading, cluttered background

Settings: 1024 x 1024, CFG 6, 30 steps, fixed seed

Avoid asking for captions in the image: Stable Diffusion handles text poorly. Add them afterward in your word processor. And check the illustration's content, because a generated diagram can be wrong while looking authoritative.

For SD 3.5, write in full sentences instead:

A close-up photograph of [SUBJECT] in [PLACE], taken in the early morning. The light comes from the left and leaves the background slightly out of focus. In the foreground there is [IMPORTANT DETAIL]. The overall palette is [2 OR 3 COLORS], with a calm and natural feel.

Start without a negative prompt and add one only if a specific flaw persists: the official examples for this model leave that field empty.

LoRAs and embeddings: what they change in your prompt

A LoRA is a small file that adds a style, a character, or a concept to the model. In AUTOMATIC1111, it's called in the prompt with <lora:file_name:0.8>, where the number sets its strength. In ComfyUI, it's loaded with a dedicated node. Most LoRAs also need a trigger word, listed on their page: without it, the effect may be weak or nonexistent.

An embedding, or textual inversion, behaves like a custom word. In ComfyUI, it's written embedding:name, and with a weight (embedding:name:1.2). Its place in the prompt matters, as with any ordinary word.

Always check three things before installing one: its compatibility with your model family, since an SD 1.5 LoRA doesn't work on SDXL; its license, especially for commercial use; and its origin. A LoRA that imitates a living artist or a real person raises questions covered in our article on commercial use.

Make your images reproducible

A Stable Diffusion prompt without its settings is incomplete. To reproduce an image, you need the exact model, the positive and negative prompts, the seed, the CFG, the number of steps, the sampler, the size, and any LoRAs with their strength. Even with all that, a change in interface version or graphics card can create small differences.

The tools help: AUTOMATIC1111 writes the parameters into the PNG's metadata, and ComfyUI saves the workflow there. Be careful, though: platforms that recompress your images often strip this data, so keep your originals.

Keep a small notebook too, with the settings and a reference image for each prompt that works. A simple file is enough, and our article on reusable prompt libraries suggests a way to organize it. This rigor becomes essential if you publish your prompts: in YouPrompt's AI images section, the buyer sees a real result before purchasing, and they can only get close to it if you state the model and the settings you tested.

To get started, pick an image you like among your generations, fix its seed, and compare three CFG values without touching anything else. Write down the winner next to the model name: that'll be your starting setting for everything else.