Algorithm evaluation demonstrates improved style transfer fidelity and semantic preservation in diffusion models, suggesting dynamic layer routing mitigates content leakage.
Key Points
To prevent content leakage and semantic distortion during style transfer in diffusion models by developing a dynamic routing architecture.
Designed DynaStyle, an end-to-end framework featuring a dynamic gating mechanism with 0.5M parameter overhead for layer-adaptive conditioning.
Implemented stage-aware feature fusion through learned attention reweighting across transformer layers during diffusion denoising, ensuring compatibility with custom base-model variants.
Achieved state-of-the-art performance across diverse content and style domains.
Successfully disentangled style representations from content to eliminate semantic distortion while preserving target content structure and stylistic coherence.