Negative Prompt: Crucial for SD. List what you don’t want (e.g., “Blurry, distorted hands, extra limbs, low resolution”).
SHORTCUTS:
Attention/Weighting: Use parentheses to emphasize specific words (photorealistic:1.5) makes it much more realistic).
Prompt Blending: Use the [from:to:when] syntax to switch subjects mid-gen [cat:dog:0.3] starts as cat & finishes as dog.
Shorten & Test: Start with a simple prompt and add only 1–2 keywords at a time to see their specific impact.
Styles Dropdown: In tools like AUTOMATIC1111, use the “Styles” menu to save and quickly apply complex prompt/negative prompt combinations.
CHEAT CODES:
Sampler Selection: Choose Euler A for speed and soft results, or DPM++ 2M Karras for high-detail, realistic textures.
CFG Scale: Usually kept between 7–9. Higher values (15+) follow your prompt strictly but can “burn” the image; lower values (3–5) give the AI more creative freedom.
Steps: 20–30 steps are usually enough for modern models (like SDXL); going higher (50+) rarely improves the image but increases generation time.
Resolution: Stick to the native resolutions of your model for the best quality (e.g., 512×512 for SD 1.5, or 1024×1024 for SDXL).
ADVANCED:
ControlNet: Use this to guide the AI with a specific structure (like a pose, a depth map, or a sketch) instead of just words.
LoRA (Low-Rank Adaptation): Small, downloadable “add-ons” that teach SD a very specific character, object, or artistic style (e.g., “Studio Ghibli LoRA”).
Img2Img: Upload your own image and use a text prompt to “transform” it into a new style while keeping the original composition.
Inpainting: Use a brush to “paint over” a specific part of an image you want to change, like swapping a character’s shirt or fixing a background error.
