What’s Wrong with AI Generation?

Text-to-image diffusion models struggle with compositional accuracy — failing to count objects correctly, respect spatial layouts, and avoid mixing object attributes. A training-free, plug-and-play method called 'attention refocusing' is proposed to fix this. It uses a large language model to generate a spatial layout (bounding boxes) from the text prompt, then guides both cross-attention and self-attention maps during the diffusion sampling process using a differentiable loss function. This approach improves object count accuracy, spatial composition, color fidelity, and reduces object hallucination across multiple existing models without requiring retraining.