10 promos for Stable Diffusion that will help you create original pictures

How to write queries for Stable Diffusion

When working with the Stable Diffusion neural network and other similar systems, the most successful results can be obtained if you create detailed instructions. Here's how to do it right.

Include all important details

First, you need to come up with the purpose of the request and the approximate result, in what style it should be presented and what cannot be depicted in the picture.

The correct piece must contain the subject or object that we want to see in the image. You need to specify what the subject does, its location, and then any additional keywords. A complex request may also include a detailed description of the type of image: drawing, realistic photograph, landscape. Often this property can be activated in the service settings.

The closer to the beginning of the request it is, the more power it has in generating the image. For each element, you can specify the importance or “weight” manually. This is usually written after a colon without a space, for example: weight: 1.5, weight: 0.4.

In detailed queries, it is worth using additional words that will help the neural network select the most successful option. Describe all the characteristics that you can think of for a specific task.

You can find the right keywords by experimenting on your own or using a product designer. In the second case, you can immediately take into account most of the available parameters and modifiers of the final result.

Register unnecessary elements and characteristics

A negative prompt can also help when composing a request. This is a list of all the properties that should not be in the image. Often it indicates points that depend on the context.

It may be necessary to exclude deformed body parts, a second person in the frame, blur, poor quality and other similar things. Negative promt will help you collect frames with fewer artifacts. But you don’t have to make huge lists of negatives—basic keywords are enough.

Take into account the main parameters in the request settings

The neural network settings always contain several important parameters: generation steps, seed, resolution, sampler.

• Steps - indicates the number of steps the neural network will take while generating the image. The more there are, the higher the quality of the result usually is.
• Seed - this is the reference point on the basis of which the system generates a picture. Initially, a “random” parameter from billions of options is specified. This helps to get different results for the same query. If you use a specific value, then when you change the prompt, the contents of the image will remain approximately the same.
• Sampler - means processes that influence the final result. For simple shots, 5-10 steps are enough, while for others you will need 60-90.
• Resolution is the size of the image. The Stable Diffusion XL model generates a maximum of 2,048 x 2,048 pixels.

For a beginner, it’s better to start with a promt constructor. The most popular option with a large number of settings is Promptomania. It is suitable for various neural networks, including a section for Stable Diffusion.

What commands might be useful?

At the moment, the latest neural network model is Stable Diffusion XL (SDXL). It provides high quality images and has significant improvements over previous versions. Options that work in SDXL may differ from earlier models.

To test requests in this case, we used the online service DreamStudio based on Stable Diffusion XL 1.0. Next, let's look at examples of promts for generating pictures in various styles and forms.

1. For photorealistic images

Stable Diffusion XL can generate images in a wide variety of styles. One of the most important is photorealism. It's unlikely that you'll get a perfect result the first time, but after a series of attempts you can generate impressive images.

Examples of possible queries:

• Metal blue pen made by Parker, fountain pen, reflective, positioned on the work table. The pen is on top of the paper notebook with black and gold cover design. Sunny, bright image. Huge copy space on top of the image, negative space, hd, 8k, blurry dreamy background, beautiful pictures.
• Young woman in fabulous dress, photorealism, dark hair, centered in the frame, symmetrical face, 50 mm lens, facing camera, dark background, natural light, out of focus mountains in the background, ultra-details, 16K.

Suitable negative details for queries: (deformed, distorted, disfigured: 1.3), poorly drawn, out of frame, watermark, signature, cut off, low contrast, underexposed, overexposed, bad art, beginner, amateur.

2. For animated pictures

The neural network allows you to create animated images of any complexity, but to get a good result you will have to tinker with the query assembly.

Example requests:

• Man with black space helmet, long dark coat, body portrait, red light on his chest, fashion show concept boots, sharp focus, black scarf, windy, highly detailed, red eyes, digital painting, illustration art.
• Puss in boots sitting in a restaurant eating macaroni and cheese watching a singer performance, unreal engine, cozy indoor lighting, artstation, detailed, digital painting, cinematic, illustration art.
• An illustration of a big dragon breathing fire, small people around him praying. Intense and powerful fiery breath, volumetric. Emphasize its piercing gaze, sharp teeth and powerful stance. Artstation, detailed, digital painting, cinematic.

Negative details for requests: (deformed, distorted, disfigured: 1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers: 1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, bad anatomy, watermark, signature, cut off, low contrast, underexposed, overexposed, bad art, beginner, amateur.

3. For artistic landscapes

Stable Diffusion generates excellent landscapes, for example, an autumn forest in the rays of the setting sun, fantastic locks, beaches or futuristic cities.

Example requests:

• Big forest with yellow, green and red leaves at sunset, a small town with a church and a radio tower in the background, fantasy, professional canvas, triadic colors, deep color, volumetric lighting, shading with dark edges, richly detailed, matte background, style octane render.
• The environment in a fantasy world of Tolkien, gothic big castle, ultra‑detailed, daytime, movie concept art, epic, matte painting, mist, volumetric light, 16k, artstation.
• A painting of a rocky beach with grand waves crashing on it and a thunder in the background with a mansion on the hill, extremely detailed oil painting, volumetric light, 16k, artstation.
• Futuristic highrise, flash light, epic composition, cinematic, landscape illustration, detailed painting, 8k post processing, unreal engine, artstation.

Negative details for queries: (deformed, distorted, disfigured: 1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers: 1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, people, human, signature, watermark, sign.

4. For portraits

Stable Diffusion XL does a good job of generating portraits, but super-realism is still difficult to achieve. But the neural network is useful for testing options for real shooting or, for example, comics.

Example requests:

• A young female, green ribbons in her hair, sitting outside of the cafe in Paris, green eyes, wearing a stylish dress, side light, photorealistic
• Gentleman in a black suit, ultra realistic, cinematic shot, highest quality, (detailed face and eyes: 1.1), short hair, inside an old mansion, subsurface scattering, (detailed skin: 1.1), shiny skin, professional photo, volumetric lighting
• Stylish woman, street photography shot, long blonde hair, capturing her reflection in a bookstore glass frontdoor, upper body framing, on a Parisian street, neon lighting, shot from a low angle, (bokeh effect: 1.2), ultra realistic

Negative details for queries: (deformed, distorted, disfigured: 1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers: 1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, bad anatomy, watermark, signature, cut off, low contrast, underexposed, overexposed, bad art, beginner, amateur, distorted face.

5. To generate characters

With Stable Diffusion XL you can create beautiful and funny shots of famous characters. You can combine several different styles at the same time.

Example requests:

• A cinematic shot of a Batman, filmed on Red V‑Raptor 8K camera, 50 mm, f/1.4, film director David Fincher, chiaroscuro, full‑body, futuristic cyberpunk costume, magic realism.
• Detailed 8k image of One‑Punch Man seated on a metal throne amidst Japanese architecture at night. Illustration, realistic, anime, volumetric lighting.
• A picture of Vault Boy from Fallout as a gladiator in a big fighting arena, haze air, illustration.

Negative details for queries: (deformed, distorted, disfigured: 1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers: 1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, people, human, signature, watermark, sign.

6. To create architectural objects

SDXL is useful for designers. For example, it can be used to create conceptual images of buildings in any style. The neural network often makes mistakes in the lines and shapes of complex objects; you will have to repeat the generation process until a successful result is obtained.

Example requests:

• Architectural design plan of a guest house shot from outside, large steel windows with a view of the city, high ceiling, high quality, 8k, ultra realism, high resolution photography
• Exterior shot of a villa in Mykonos, with sea view, white and blue colors mood, volumetric lighting, high quality, 8k, ultra realism, high resolution photography
• Rustic house in the spanish style, made of stone, with white and blue decoration. volumetric lighting, high quality, 8k, ultra realism, high resolution photography

Negative details for queries: (deformed, distorted, disfigured: 1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers: 1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, people, human, signature, watermark, sign.

7. For interior design

The neural network is also suitable for preparing the interior design of premises. Use SDXL if you want to find unusual options for your home design or apartments.

Example requests:

• A futuristic living room with a reflective ceiling, high metal windows, (futuristic decor: 1.2), future cityscape at night in the distance, high quality, 8k, ultra realism
• Kitchen in the style of Sou Fujimoto, wide angle view of a vibrant city in the windows, interior architecture, rendered in octane, photorealistic, highly detailed
• Interior design project with modern organic style, room with sofa, carpet, magazine table, chandelier and armchair, neutral colors, bright environment, realistic, V‑Ray renderer, highly detailed

Negative details for requests: (deformed, distorted: 1.3), poorly drawn, disgusting, blurry, people, human, signature, watermark, sign.

8. To create images of clothing

SDXL offers good footage with clothes, but you will have to experiment with queries and repeat the process many times to get successful pictures.

Example requests:

• Techwear clothes and fashion look, feet to head, golden ratio, futuristic, highly detailed, Errolson Hugh, Yohji Yamamoto, Nike ACG, hyper maximalist.
• High fashion dress and fashion look, feet to head, golden ratio, highly detailed, pastel colors, Miu Miu, Prada, luxury, high quality, 8k, ultra realism.

Negative details for queries: (deformed, distorted, disfigured: 1.3), bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers: 1.4), disconnected limbs, mutated, blurry.

9. To create logos

The neural network will also come in handy during the formation of the organization’s identity. With SDXL you can test different logo ideas in any style. Be prepared for constant adjustments to requests.

Example requests:

• A minimalist logo featuring a retrofuturistic UFO. The logo showcases UFO flying over skyscrapers and glowing city lights. The color palette is a mix of dark and intense colors, including shades of black and red. Illustration, vector art, logo design.
• Logo for a technology company, white and blue, a seagull flying, symmetrical, illustration, vector art, logo design.

Negative details for requests: (deformed, distorted, disfigured: 1.3), poorly drawn, disgusting, blurry, people, human, signature, watermark.

10. For optical illusions

Stable Diffusion allows you to generate optical illusions - images based on references or pictures with words “sewn” into them. They are created using the ControlNet neural network plugin. It allows you to obtain pictures based on other images, preserving elements from the original frames.

The neural network creates a picture at the user’s request and takes into account the proposed reference. For example, the landscape is combined with the form stars or the grass forms a word that is visible when moving away from the image.

To use Stable Diffusion with the ContolNet plugin, you will need a powerful computer. As an alternative, you can use online services with the same capabilities. You can read more about the available tools and how to work with them in article Life hacker.