Prompt structure that works: SD 1.5 versus SDXL

The two model families read prompts differently, and most "the model ignores me" complaints are an SD 1.5 prompt fed to SDXL or the other way round. SD 1.5 (and its fine-tunes, such as Dreamshaper 8) SD 1.5 was trained on short, tag-like captions. It responds best to a compact, ordered list: subject, action, setting, lighting, medium or style, camera or lens terms Example: portrait of an old fisherman, mending a net, harbour at dawn, soft side light, oil painting, 85mm Put the most important thing first. Words early in the prompt carry more weight than words at the end, and anything past the token limit is handled in a separate chunk (see the 77-token thread). SDXL SDXL uses two text encoders and was trained on longer, more descriptive captions. A plain sentence often beats a tag pile: An old fisherman mends a net on a stone harbour wall at dawn, soft light from the left, painted in oils. SDXL also needs fewer "quality" tags. Strings like masterpiece, best quality, 8k came from anime-model habits on SD 1.5 and do little on base SDXL. Resolution matters as much as wording SD 1.5 was trained at 512x512; SDXL at about one megapixel (1024x1024, 832x1216, 1216x832 and similar). Render SD 1.5 at 1024 directly and you get doubled heads and repeated limbs, because the model tiles its idea of a 512 picture. Render at native size and upscale afterwards. Reproduce it Same prompt, same seed, both phrasings, one model. Change one thing at a time and keep a note of the seed. A post without the seed is a picture; a post with the seed is a method. Background: https://wiki.soapbox.community/wiki/Generative_Image_AI_and_ComfyUI