# Prompt components A good prompt is less about poetry and more about engineering. The diffusion model doesn't "understand" language like a human; it maps words to visual concepts (embeddings). To communicate effectively, you need a structured approach. ^overview Most high-quality images follow a consistent "slot-based" structure. Thinking in slots prevents you from rambling and ensures the model has the necessary information to render the image correctly. ### The 6 Core Components 1. **Subject**: *What* is in the image? Be specific. - "A cat" is weak; "A fluffy white Persian cat" is strong. ^subject-core-component 2. **Medium**: *How* was it made? - (e.g., "Photograph," "Oil painting," "3D render," "Graphite croquis"). - If you omit this, the model usually defaults to a generic "digital art" look or something else if the model is already specialized. ^medium-core-component 3. **Style**: *Who* made it? - (e.g., "Style of Greg Rutkowski," "Studio Ghibli," "Cyberpunk 2077"). - You can mix styles ("50% Van Gogh, 50% Pixar"). ^style-core-component 4. **Composition**: *How* is it framed? - (e.g., "Wide angle," "Macro shot," "Isometric view," "Rule of thirds"). - This constrains the camera. ^composition-core-component 5. **Lighting**: *What* is the mood? - (e.g., "Volumetric lighting," "Golden hour," "Cinematic lighting," "Flat lighting"). - Light defines the dimensionality. ^lighting-core-component 6. **Quality**: *How* polished is it? - (e.g., "4k," "8k," "Masterpiece," "Best quality"). - These are "magic words" that bias the model towards high-aesthetic training data. ^quality-core-component ### Token Ordering vs. "Word Salad" * The model pays the most attention to the **start** of the prompt. * **Front-load the subject and medium**: `[Subject] [Medium] ...` * Move style and quality modifiers to the end. * Avoid "word salad"—long lists of loose adjectives often confuse the cross-attention mechanism, leading to "concept bleeding" (e.g., asking for a "blue hat" and "red shoes" might result in red hats and blue shoes). ^anatomy-components