Few-Shot Atypical Layout-to-Image Generation via Disentangled Semantics
Researchers propose a representation-driven framework for layout-to-image (L2I) generation that addresses representation fragmentation in few-shot atypical settings. The method disentangles semantics from primitives using Semantic Anchoring for stable identity, Primitive Imbuing for robust local detail, and Conceptual Steering for foreground consistency. Experiments show improvements over state-of-the-art L2I methods in the 5-shot regime.
Key facts
- arXiv:2605.31266v1
- Layout-to-image task enables fine-grained control via object categories and spatial layouts
- Existing L2I methods fail under few-shot atypical settings due to representation fragmentation
- Granularity mismatch entangles semantic identity with visual details
- Semantic Anchoring aggregates categorical semantics into anchors
- Primitive Imbuing models recomposable primitives
- Conceptual Steering uses saliency-aware objective
- Consistent improvements in 5-shot regime over state-of-the-art
Entities
Institutions
- arXiv