DiFuse: Fusing Native Diffusion Representations for Spatial Reasoning in Unified Multimodal Models
Unified multimodal models (UMMs) jointly support visual understanding and image generation, yet information flow between these two capabilities remains largely asymmetric: understanding representations routinely condition generation, whereas generative representations are rarely reused for multimodal reasoning.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv