EVENT DOSSIER
To address the issues of poor identity consistency and weak context control in the single-reference image diffusion model used for generating personalized stickers, researchers proposed the SEAL module. As a plug-and-play adapter that does not require modifications to the U-Net backbone network, this module effectively solves the problem of overfitting by introducing three core components: semantic-guided spatial attention loss, split-and-merge token strategy, and structural-aware layer constraints. To support attribute-level control, the team also released StickerBench, a large-scale sticker dataset containing six categories: appearance, emotion, action, composition, style, and background. Experimental validation shows that SEAL significantly improves context controllability while maintaining identity consistency; the relevant code and dataset are available for public use.
研究人员提出名为 SEAL 的模块，旨在解决单参考图像扩散模型生成中贴纸个性化易过拟合的问题。SEAL 是一个无需修改 U-Net 骨干网络的即插即用适配器，通过语义引导空间注意力损失、分合 Token 策略及结构感知层限制三个组件工作。为支持属性级控制，团队发布了包含六类标签（外观、情绪、动作、构图、风格、背景）的大规模贴纸数据集 StickerBench。实验表明，SEAL 在保持身份一致性的同时维持了上下文可控性，相关代码与数据集将公开。