RoPE

FreeStyle: Dual-Reference Image Generation via Community LoRA Mining with Leakage Suppression

FreeStyle is a new framework for style-content dual-reference image generation that mines community LoRA models as compositional anchors to construct large-scale training triplets. The approach uses a two-stage curriculum with attention-level enrichment constraints and frequency-aware RoPE modulation to suppress semantic leakage between style and content references. The authors also introduce a benchmark with novel metrics including a style-invariant Content Alignment Score and a VLM-based Rejection Score for evaluating leakage suppression.