SPARGen: Unifying Spatial Perception and Reasoning through Native Multimodal Generation
SPARGen treats reconstruction, correspondence, and spatial reasoning as one generation problem.
The paper presents a native multimodal framework that uses instructions to produce both tokenized structured outputs and dense image-aligned geometric fields. Its aim is to let supervision from different spatial tasks shape shared representations, instead of relying on separate task-specific models or external geometry modules. The authors report competitive benchmark results across 3D reconstruction, dense correspondence, and spatial reasoning within the same system. HF Daily Papers' note
The paper presents a native multimodal framework that uses instructions to produce both tokenized structured outputs and dense image-aligned geometric fields. Its aim is to let supervision from different spatial tasks shape shared representations, instead of relying on separate task-specific models or external geometry modules. The authors report competitive benchmark results across 3D reconstruction, dense correspondence, and spatial reasoning within the same system. HF Daily Papers' note
score 5