Shanghai AI Lab's Intern Lumina U2 Unifies Image, Video and 3D in a 16B Multi-Codebook Diffusion LLM

Intern Lumina U2, posted to arXiv on October 16, is a 16B / 1.4B-active masked diffusion LLM that models images, videos and 3D assets as discrete multi-codebook tokens over a shared backbone with text. The team reports first-place results on 6 of 10 image-understanding benchmarks among unified models, GenEval 89.3 on image generation, and a 4.45x end-to-end latency speedup at 1024x1024 via plateau-Halton sampling with V-verify. Training spanned 1,024 Ascend 910B NPUs across 1.05T tokens in five progressive stages.