I sank hours into Uni-1 from @LumaLabsAI, and yeah—this feels like a line being crossed. Most image models are basically “text in, pixels out.” Uni-1 pauses and seems to work out the scene first, then renders it. Look closely: the glass bending light, reflections behaving the way they should, surfaces catching illumination, clean depth separation. This isn’t some clever prompt gymnastics carrying the result. It reads like the model is actually reasoning. That’s the real change here: less about phrasing the perfect ask, more about the system actually understanding what you meant. At the moment, I’d rank it above NB Pro and GPT Image 1.5.