How to get the best results from Stable Diffusion 3

This post introduces Stable Diffusion 3 Medium, detailing its three text encoders (CLIP and T5-XXL) and how different versions of the model are packaged based on VRAM requirements. It highlights SD3's improved prompt adherence with significantly longer prompts (up to 10,000 characters) and advises against using negative prompts, explaining their effect as introducing noise rather than removing unwanted elements. The post provides specific prompting techniques, including using plain English and being descriptive, and suggests experimenting with different prompts for each text encoder. It also offers recommended settings for image generation, including optimal resolutions for various aspect ratios (around 1 megapixel), number of steps (28 recommended), guidance scale (3.5-4.5), sampler (`dpmpp_2m`), and scheduler (`sgm_uniform`). The impact of increasing steps on image coherence and potential subject changes is discussed, along with the consequences of generating images at resolutions larger or smaller than optimal.