时序证据双门控青花瓷风格迁移

Time-Evidence Dual Gating for Blue-and-White Porcelain Style Transfer

摘要: LoRA微调技术已广泛应用于稳定扩散(Stable Diffusion)文生图模型，但在青花瓷高精度图像生成中，难以同时保证纹饰风格学习与器型结构完整性，常出现形变。本文提出一种适用于Transformer架构、无需额外训练的LoRA风格迁移方法：将LoRA注入过程解耦为时间门控与提示词证据门控，并通过乘法融合动态调节风格强度——在结构敏感阶段抑制扰动，在细节生成阶段增强纹饰表现。实验证明，相比经开发集调参后的恒定强度对比，LPIPS相对下降52.1%，SSIM相对提升31.1%，MS-SSIM、PSNR与DISTS亦呈一致改善，有效实现了结构保持下的高质量风格迁移。

Abstract: LoRA (Low-Rank Adaptation) fine-tuning has been widely adopted in Stable Diffusion for text-to-image generation. However, when applied to high-fidelity image synthesis of blue-and-white porcelain, it often struggles to simultaneously preserve structural integrity of vessel forms and accurately capture intricate decorative styles, frequently resulting in geometric distortions. To address this challenge, we propose a novel LoRA-based style transfer method compatible with Transformer architectures that requires no additional training. Our approach decouples the LoRA injection process into two dynamic gating mechanisms: a time-aware gate and a prompt-evidence gate. These gates are multiplicatively fused to adaptively modulate style intensity—suppressing stylistic perturbations during structure-sensitive denoising steps while enhancing decorative detail rendering in later stages. Experimental results demonstrate significant improvements over a baseline using constant, development-set-tuned LoRA strength: LPIPS decreases by 52.1%, SSIM increases by 31.1%, and consistent gains are observed across MS-SSIM, PSNR, and DISTS metrics. This confirms the effectiveness of our method in achieving high-quality style transfer while faithfully preserving structural fidelity.