Delta-K: Boosting Multi-Instance Generation via Cross-Attention Augmentation

A team of researchers led by Zitong Wang has introduced Delta-K, a novel inference framework designed to solve a persistent flaw in text-to-image diffusion models: concept omission. When generating complex scenes with multiple objects, models like Stable Diffusion or DALL-E 3 often fail to include all requested elements. Existing fixes, which simply rescale attention maps, tend to introduce noise. Delta-K takes a fundamentally different approach by operating directly in the cross-attention Key space. Using a vision-language model, it extracts a differential key (ΔK) that encodes the semantic signature of missing concepts and injects this signal during the early, critical planning stage of the image synthesis process.

This injection is governed by a dynamic scheduling mechanism that grounds diffuse noise into stable structural anchors for the missing objects, without disrupting elements that are already being generated correctly. The method is both backbone-agnostic and plug-and-play, meaning it works with modern DiT-based architectures and older U-Net models alike, requiring no additional training, spatial masks, or modifications to the underlying model. Extensive experiments confirm that Delta-K consistently improves the compositional alignment and fidelity of generated images, ensuring that complex prompts with multiple subjects are rendered completely and coherently for the first time.