Envisioning Beyond the Few: Disentangled Semantics and Primitives for Few-Shot Atypical Layout-to-Image Generation

Researchers propose a novel framework for layout-to-image generation that improves visual quality in few-shot learning scenarios by disentangling semantic identity from visual details. The method uses semantic anchoring and primitive imbuing to address representation fragmentation, enabling more coherent image synthesis from sparse training data.

This research addresses a fundamental challenge in computer vision: generating high-quality images from layout specifications when training data is severely limited. Traditional layout-to-image methods struggle with few-shot adaptation because they conflate categorical semantics with fine-grained visual details, leading to fragmented and distorted outputs. The proposed disentanglement approach separates these concerns, allowing models to maintain stable object identity while flexibly modeling local details.

The framework introduces three key innovations: Semantic Anchoring aggregates category-level information into stable anchors that preserve object identity across variations, Primitive Imbuing learns recomposable visual components for robust detail generation, and Conceptual Steering applies saliency-aware optimization to prioritize foreground consistency. This architectural philosophy reflects broader trends in deep learning toward modular, interpretable representations that generalize better under limited data.

For the AI research community, this work has direct implications for practical applications requiring few-shot image generation—design tools, content creation platforms, and data augmentation pipelines. The public code release accelerates adoption and enables downstream innovations. The consistent improvements across diverse atypical domains suggest the approach generalizes beyond narrow use cases, validating the disentanglement principle as a robust solution to representation fragmentation.

Looking forward, similar decomposition strategies may benefit other generative tasks facing few-shot constraints, particularly in multimodal systems where semantic and perceptual information must be carefully balanced.

• →A new framework disentangles semantic identity from visual primitives to improve few-shot layout-to-image generation.
• →Semantic Anchoring and Primitive Imbuing work together to maintain stable object categories while enabling flexible detail modeling.
• →Saliency-aware optimization prioritizes foreground consistency to preserve semantic correctness during few-shot adaptation.
• →Experimental results show consistent improvements over state-of-the-art methods across diverse atypical domains in 5-shot regimes.
• →Open-source code availability enables rapid adoption and further research into disentangled representation learning.