Luma Uni-1 debuts, thinks in pixels

Uni-1 looks like Luma's attempt to make image generation behave more like reasoning than roulette. That's the right battle to pick: useful creative tools win on editability and consistency, not on one-off pretty samples. The report says Uni-1 is state of the art on RISEBench and ranks first in human-preference Elo for overall, style/editing, and reference-based generation, while also placing second for text-to-image. The pitch maps to real pain points: prompt decomposition, source-grounded control, multi-turn refinement, and better handling of text, memes, manga, and sketches. Community chatter is already comparing it favorably with Nano Banana 2, GPT Image 1.5, and Midjourney on complex prompts. API access is still waitlisted, so this reads more like a strategic preview than a production-ready platform.