A new arXiv paper introduces CoBit, a continuous bitstream diffusion approach originally designed for language modeling. The authors report that CoBit now serves as a unified framework for multimodal generation, covering both joint and conditional tasks.
According to the abstract, the method uses the same training and sampling protocol across modalities, eliminating the need for task-specific adaptations. This suggests a path toward simpler, more consistent multimodal systems.
The abstract does not provide experimental details or comparisons, so the claims are limited to the framework's conceptual unification. Further reading of the full paper would be needed to assess performance.