Delta-K: Boosting Multi-Instance Generation via Cross-Attention Augmentation
This addresses concept omission in text-to-image synthesis for users of diffusion models, offering a plug-and-play solution that is incremental but effective.
The paper tackles the problem of concept omission in diffusion models when generating multi-instance scenes by proposing Delta-K, a training-free framework that injects semantic signatures of missing concepts into cross-attention keys, resulting in improved compositional alignment across different architectures without additional training.
While Diffusion Models excel in text-to-image synthesis, they often suffer from concept omission when synthesizing complex multi-instance scenes. Existing training-free methods attempt to resolve this by rescaling attention maps, which merely exacerbates unstructured noise without establishing coherent semantic representations. To address this, we propose Delta-K, a backbone-agnostic and plug-and-play inference framework that tackles omission by operating directly in the shared cross-attention Key space. Specifically, with Vision-language model, we extract a differential key $ΔK$ that encodes the semantic signature of missing concepts. This signal is then injected during the early semantic planning stage of the diffusion process. Governed by a dynamically optimized scheduling mechanism, Delta-K grounds diffuse noise into stable structural anchors while preserving existing concepts. Extensive experiments demonstrate the generality of our approach: Delta-K consistently improves compositional alignment across both modern DiT models and classical U-Net architectures, without requiring spatial masks, additional training, or architectural modifications.