360CVGroup/RelaCtrl

Implements relevance-guided control mechanisms that selectively inject spatial conditioning into diffusion transformer layers, enabling efficient multi-modal control (canny edge maps, style guidance) with minimal parameter overhead (45M adapters). Integrates with PixArt-α for text-to-image generation, leveraging T5 text encoding and diffusion-based latent space manipulation to achieve fine-grained spatial control without full model retraining.

383 stars. No commits in the last 6 months.