Composition for generated images: value first, then thirds

The model will happily render a technically perfect image with no focal point. Composition is still your job. Value before colour Value is lightness and darkness. A picture that reads in greyscale reads in colour. Desaturate your output (in any editor) or squint at it: if the subject does not separate from the background, no prompt about colour will fix it. How to steer value: - Lighting words with a direction: rim light, light from the left, single window light, low key. - Contrast between subject and ground: "dark figure against a bright sky". - img2img from a rough value sketch: paint three or four grey shapes where you want them, use denoise 0.6-0.75, and the model keeps your light pattern. Placement - The rule of thirds is a starting point, not a law: put the eye line or horizon on a third rather than dead centre. - The model centres subjects by default, because training images were often centre-cropped. To move the subject, use a pose or depth ControlNet, or render wider and crop. - Aspect ratio is composition. Portrait 2:3 for figures, landscape 3:2 or 16:9 for scenes; on SD 1.5 that means about 512x768 or 768x512. Method 1. Make a 3-value thumbnail in any paint tool. 2. img2img it with your prompt at denoise 0.65. 3. Desaturate the result. Does it still read? If not, adjust the thumbnail, not the prompt. Further reading on making a picture for a post: https://wiki.soapbox.community/wiki/Making_an_Image_for_Your_Post