Zifeng Cheng
Publications
TraceRouter: Robust Safety for Large Foundation Models via Path-Level Intervention
Despite their capabilities, large foundation models (LFMs) remain susceptible to adversarial manipulation. Current defenses predominantly rely on the "locality hypothesis", suppressing isolated neurons or features. However, harmful semantics act as distributed, cross-layer circuits, rendering such localized interventions brittle and detrimental to utility. To bridge this gap, we propose \textbf{TraceRouter}, a path-level framework that traces and disconnects the causal propagation circuits of illicit semantics. TraceRouter operates in three stages: (1) it pinpoints a sensitive onset layer by analyzing attention divergence; (2) it leverages sparse autoencoders (SAEs) and differential activation analysis to disentangle and isolate malicious features; and (3) it maps these features to downstream causal pathways via feature influence scores (FIS) derived from zero-out interventions. By selectively suppressing these causal chains, TraceRouter physically severs the flow of harmful information while leaving orthogonal computation routes intact. Extensive experiments demonstrate that TraceRouter significantly outperforms state-of-the-art baselines, achieving a superior trade-off between adversarial robustness and general utility. Our code will be publicly released. WARNING: This paper contains unsafe model responses.
FashionPose: Unified Text-Driven Fashion Synthesis with Joint Geometric and Photometric Control
Realistic and controllable garment synthesis is essential for fashion e-commerce, yet it demands precise coordination between human pose geometry and environmental photometry. Conventional pose-guided frameworks suffer from two fundamental limitations: they rely heavily on predefined skeletons from off-the-shelf estimators, restricting semantic flexibility; and they predominantly focus on studio-like generation under neutral lighting, failing to reconcile geometric configurations with complex, scene-specific illumination described in natural language. To bridge this gap, we propose FashionPose, a cascaded architecture that reconciles geometric and photometric control within a unified language-driven interface. Unlike conventional frameworks, our framework employs a decoupled yet synergistic strategy: (1) a bidirectional contrastive alignment mechanism that grounds textual semantics into an explicit geometric manifold, enabling template-free pose generation; (2) an identity-anchored synthesis module that translates these geometric priors into high-fidelity imagery while preserving fine-grained appearance; and (3) a prompt-conditioned relighting module that leverages the generated pose as a spatial anchor to achieve environment-aware shading. This hierarchical design effectively transforms high-level instructions into consistent visual representations, ensuring both structural precision and atmospheric harmony. To facilitate this paradigm, we construct PoseCap, a dataset with over 40,000 caption-keypoint pairs. Extensive experiments demonstrate that FashionPose outperforms existing benchmarks in pose accuracy and physical realism, providing a robust solution for personalized, scene-aware virtual fashion displays.