Closeup photo of a woman wearing a white dress on grass These are generated using the "3 Clip Approach" proposed by Matt3o youtube.com/watch?v=OrST6Nq1NUg TLDR: Various ways to "hack" SD3 when you see bad generation: 1. Use non-standard resolution such as 1042x1042 (not multiple of 64). 2. Avoid words such as "lying". Or use the 2 CLIP + T5 ClipTextEncoderSD3 node, but only use "lying" in clip_g. 3. Use "random noise" such as "aaaaa aaaaaaa aaaaa" in your negative prompt.