XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiTModulation

Achieving fine-grained control over subject identity and semantic attributes (pose, style, lighting) in text-to-image generation, particularly for multiple subjects, often undermines the editability and coherence of Diffusion Transformers (DiTs).