Stable Diffusion Prompt Book

When AI image creators are still troubled by the “fuzzy expression of prompt words leading to generation bias, difficult to control the stylized effect, and no direction for parameter settings”, the Stable Diffusion Prompt Book (hereinafter referred to as the “Prompt Book”) launched by OpenArt platform becomes the “creation manual” for Stable Diffusion users with the three core advantages of “free open source, systematic guidance, and case-based teaching”. Prompt Book” (hereinafter referred to as “Prompt Book”), with the three core advantages of “free open source, systematic guidance, case-based teaching”, has become the “creation manual” for Stable Diffusion users. As an open source manual written by Mohamad Diab (PublicPrompts) and other industry experts, Prompt Book builds a complete knowledge system of “theory-case-practice” from basic syntax to advanced techniques, and from style thesaurus to parameter suggestions, allowing even zero-basic users to master the “Generation” of Stable Diffusion within one hour. Within one hour, it masters the logic of “generating professional-grade images”, completely lowering the professional threshold of AI image creation.

Core positioning: from “fragmented techniques” to “systematic methodology”, defining the standard of Stable Diffusion cue words.

(i) Free and open source knowledge sharing: no threshold to get professional guidance

• Unlimited access: Anyone can download the PDF version from OpenArt’s official website ( https://openart.ai/promptbook) without registering or logging in, supports offline reading and printing, and as of 2025 maintains the validity of the core content as updated on November 13, 2022;
• Open source collaboration iteration: the manual content is continuously optimized based on community feedback, developers can participate in GitHub and other platforms to supplement the new style words, new cases, the formation of a virtuous cycle of “creators – experts”, for example, the subsequent community to add the “AI Generative Art Copyright Tips For example, the “AI Generative Art Copyright Tips” chapter added by the community later is originated from users’ practical feedback;
• Multi-language adaptation: In addition to the original English version, the community spontaneously translated into 8 languages, such as Chinese, Japanese, Spanish, etc. Chinese users can get the localized version through the OpenArt Chinese community, eliminating the language barrier.

(ii) Structured cue word system: make the creation logic “detachable and reproducible”.

• Four elements of the prompt structure: clear prompt word needs to contain “Subject (Subject), Style (Style), Details (Details), Parameters (Parameters)” four core elements, for example, to generate the “cyberpunk style of future cities ” prompt word, need to be disassembled as:
  • Main body: “future city, tall buildings, air transportation”;
  • Style: “Cyberpunk, neon lights, rainy wet ground”;
  • Detail: “8K resolution, cinematic light and shadow, clear detail texture”;
  • Parameters: “CFG scale 8, sampling step 20, seed value 123456”;
• weight control syntax: detailed explanation of “()” (raise the weight 1.1 times), “[]” (reduce the weight 0.9 times), “(keyword: weight value)” (customized weights) , e.g. “(neon light: 1.5)”) are used in scenarios such as reinforcing stylistic features by “(cyberpunk style: 1.2)” to avoid generating off-topic images;
• Negative Cues Guidelines: a special section on how to exclude unwanted elements (e.g., “low quality, blurry, ugly, extra limbs”) through negative cues, e.g., when generating images of people, adding “extra fingers For example, when generating character images, adding “extra fingers , missing fingers” avoids limb deformities and improves the quality of the generated images by 40%.

(iii) Case-based comparative teaching: make the difference in effect “visualized and understandable”.

• Univariate comparison experiment: for the same topic (e.g. “cats”), only change the “style words” (e.g. “ink style” vs. ” Realistic style”) or “Detailed description” (e.g. “Clear hair” vs “Blurred hair”), to show the differences in the generated images and help users to understand “keyword prioritization”;
• Error Tip Fix: Include “reverse case” (e.g. “blurry tip: ‘nice view'”) vs. “optimization tip ” (e.g. “Precise tip: ‘Alps, sunrise, oil painting style, 8K, detailed texture, golden sunlight'”), labeling the correction logic (e.g. adding “scene + style + Resolution”);
• Stylized effects disassembly: for popular styles such as “Cyberpunk”, “Steamwave”, and “Ink and Wash”, disassemble them into “Core Elements + Color Schemes + Light and shadow characteristics”, for example, “cyberpunk” needs to include “neon colors (pink / blue), holographic projection, dilapidated and technological fusion of buildings, rain reflective ground”, the user only need to press the elements of the combination of the prompt word, you can stably Generate the target style.

Core Content Framework: Five Modules to Build Stable Diffusion Creative Capabilities

（Module 1: Basic Grammar and Prompt Structure — Laying Down the Logic of Creation

• Core Syntax Rules:
  • Keyword priority: the first half of the cue word is weighted higher than the second half, e.g. “cyberpunk city, neon light” reinforces “urban subject” more than “neon light, cyberpunk city “;
  • Symbol usage specification: detail the correct usage of “comma-separated keywords” “()/[] weight control” “–ar scale parameter”, e.g. “–ar 16:9” generates widescreen images;
  • Avoid ambiguity: cues need to be concrete rather than abstract, e.g. “girl in red dress” is preferred to “pretty girl”, “8K resolution” is preferred to “HD”. “high definition”;
• Four-element prompt template:

  Provide templates that can be directly reused: “[subject description, action / state], [style keywords], [detail description (resolution, lighting, texture)], [parameter labeling (scale, CFG values)]”, for example:

  “Girl sitting by a cafe window, reading a book, realistic style, 8K resolution, warm yellow lighting, fine skin textures, –ar 3:4, CFG scale 7”.

(II) Module 2: Stylized Modifier Thesaurus — Unlocking Multiple Creative Styles

• Classic Art Styles:
  • Ink and wash style: core elements “black, white and gray dominant, brushstroke texture, white space, Chinese style architecture / figures”, example tip: “Ink and wash style landscape scroll, distant mountains and near water, fishing boat, clear brushstrokes, white space composition, -ar 2:1”;
  • Oil painting style: core element “heavy color, brushstroke texture, strong contrast of light and shadow”, example tip: “Van Gogh style sunflower field, heavy oil paint, golden sunlight, clear brushstrokes, 8K, –ar 1:1”;
• Modern Pop Style:
  • Cyberpunk: core elements “neon light (pink / blue / purple), holographic projections, high-rise billboards, rainy reflective ground, futuristic techiness”, example tip: “Cyberpunk Tokyo street, neon sign, holographic cat projections, rainy wet road, 8K, cinematic light and shadow, –ar 16:9”;
  • Steamwave: core elements “pink / blue gradient, vintage computer elements, Roman columns, palm trees, nostalgia”, example tip: “Steamwave style vintage bedroom with pink walls, old TV set, palm tree decoration, gradient light and shadow, –ar 4:3 “;
• Special Effects Style:
  • Low polygon: core element “geometric color blocks, no gradient, clean outline”, example tip: “Low polygon style fox, triangular color block stitching, orange / white dominant, clean background, –ar 1:1”;
  • Linework style: core elements “black lines, white background, detailed outline, no fill color”, example tip: “Cat linework, detailed fur outline, white background, clean lines, –ar 3:4”.

(iii) Module 3: Comparative Case Analysis – Intuitively Understanding the Impact of Cue Words

• Case 1: Comparison of accuracy of subject description
  • Fuzzy prompt: “girl” → Result: blurred face, unclear clothing;
  • Precise prompt: “20-year-old girl, brown curly hair, white dress, standing under the cherry blossom tree, smiling, hands gently clasped” → Generation result: character features are clear, scene details are complete;
  • Conclusion: Supplementing “age, hairstyle, dress, action, scene” can greatly improve the generation accuracy.
• Case 2: Comparison of style word weights
  • Normal prompt: “cyberpunk city, ink style” → Generation result: mixed styles, cyberpunk features are not obvious;
  • Weight Optimization Tip: “(Cyberpunk City: 1.2), Ink and Wash Style” → Generate result: Cyberpunk-based, fusion of ink and wash strokes, more unified style;
  • Conclusion: the style prioritization can be clarified through the weight notation to avoid “style fights”.
• Case 3: Comparison of the effect of negative cue words
  • No negative cues: “Portrait” → Generate results: there are problems of “blurred face, redundant fingers”;
  • Add negative cues: “Portrait, high quality, 8K, –no low quality, blurry, extra fingers” → generation result: clear face, normal limbs;
  • Conclusion: Negative cues are the key to “eliminating errors”, which can improve the generation quality by more than 30%.

（Module 4: Suggested Parameter Settings – Controlling Generation Details

• Key parameters are explained in detail:
  • Resolution: It is recommended that users with no basic knowledge start from 512×512, and when the performance of the device allows, it can be upgraded to 768×768 to avoid using 1024×1024 directly, which leads to a long generation time;
  • CFG scale (cue fit degree): range 1-30, recommended 7-10 — lower than 7 will generate deviation cues, higher than 10 will be stiff images;
  • Sampling steps: 20-30 steps are recommended, 20 steps can balance the speed and quality, 30 steps are suitable for the pursuit of detailed scenes;
  • Seed: -1 is randomly generated, fixed seed value can reproduce the same image, modified seed value can generate the same style variant;
• Scenario-based parameter templates:
  • Fast generation (focus on efficiency): 512×512 resolution, CFG scale 7, sampling step 20, seed value – 1;
  • Detail generation (focus on quality): 768×768 resolution, CFG scale 9, sampling step 30, seed value fixed;
  • Stylization generation (focusing on creativity): 512×768 resolution, CFG scale 8, number of sampling steps 25, add style word weight adjustment.

(v) Module 5: Advanced Techniques – From “Can Generate” to “Generate Well”

• Multi-subject synergy cue: When generating images containing multiple elements, you need to clarify the “subject relationship”, e.g. “girl sitting on the sofa, a cat next to her, the cat playing with a ball of wool, living room scene, realistic style”, to avoid confusion of elements;
• Dynamic Effect Tip: Generate a sense of dynamism through “Action Description + Dynamic Blur Words”, e.g. “Running teenager, trench coat fluttering, dynamic blur, 8K, cyberpunk city background”;
• Copyright compliance tips: add community-supplemented “copyright avoidance tips”, such as “not for commercial use, AI generated art, original character”, to avoid generating infringing content;
• Cross-platform adaptation: the content of the instruction manual is adapted to mainstream tools such as Stable Diffusion 1.5/2.1, Automatic1111 WebUI, ComfyUI, etc., and the prompts and parameters can be reused directly without adjusting the format.

Use process: five steps to master the creation of Stable Diffusion, zero basis can also get started!

（Step 1: Download the manual and learn the basic syntax (30 minutes).

1. Get resources: Visit OpenArt official website to download the PDF version of Prompt Book, and focus on reading “Module 1: Basic Syntax”;
2. Grasp the core framework: memorize the structure of “Body + Style + Details + Parameters” and understand the usage of “()/[]” weight symbols;
3. Initial practice: enter simple prompt words (such as “(cat:1.1), cyberpunk style, 8K”) in Stable Diffusion WebUI, observe the generation effect, and feel the influence of weighting symbols.

(ii) Step 2: Call the style thesaurus and determine the direction of creation (30 minutes)

1. Choose a style: Find the target style (e.g. “Ink and wash style”) in “Module 2: Stylized Modifier Thesaurus” and extract the core elements;
2. Combine prompt words: Combine prompt words according to the “Four Elements Framework”, e.g. “Ink and wash style with mountain peaks, cloudy, pine trees, brushstroke texture, white background, –ar 2:1, CFG scale 8”;
3. Generation test: enter the cue words and set the parameters as suggested in the manual, generate the first stylized image and record the deviation of the effect (e.g. “clouds are not visible enough”).

(iii) Step 3: refer to the case comparison and optimize the cue words (1 hour)

1. Find similar cases: In “Module 3: Comparative Analysis of Cases”, find cases of the same style and compare the differences of your own cue words;
2. Correct the deviation: If the generated “Ink Peaks” clouds are not obvious, refer to the case to add “(clouds: 1.2), light texture”;
3. Add negative tips: for “blurry, extra elements” problem, add “–no low quality, blurry, extra objects”, regenerate and compare the effect.

(iv) Step 4: Adjust parameters to improve the quality of generation (30 minutes)

1. Optimize the parameters: If the image details are insufficient, raise the “sampling steps from 20 to 30” and “resolution from 512×512 to 768×768” as suggested in the manual;
2. Fixed seed value: After finding a satisfactory generation result, record the seed value, fine-tune the cue words (e.g., “add flying bird element”), and generate the same style variant;
3. Batch Generation: Use WebUI’s “Batch Generation” function to input 3-5 groups of optimized prompts and filter the best results.

（Step 5: Practice advanced techniques and expand creative boundaries (ongoing)

1. Experiment with complex scenes: learn “Module 5: Advanced Skills” to generate multi-subject images (e.g. “Family dinner scene, 5 people sitting around a dining table, realistic style”);
2. Explore new styles: Combine the new style words added by the community (e.g. “AI generates art styles”) to expand the types of creations;
3. Summarize experience: Record “effective cue words + parameter combinations” to form a personal style template and enhance the efficiency of subsequent creation.