SenseNova-U1: Unified Multimodal Model Works Directly with Pixels Without VAE or Separate Encoder

arXiv·medium signal

SenseNova-U1 introduces the NEO-unify architecture that combines image understanding, generation, and editing in a single transformer without a separate visual encoder or variational autoencoder, operating directly on pixels and text. The 8B-MoT variant outperforms Qwen-Image, BAGEL, FLUX.1-Kontext, and OmniGen on editing benchmarks, while the 30B-A3B MoE model matches InternVL and Qwen2.5-VL on understanding tasks under a 32x compression ratio. Open-source release with code on GitHub.