Zero-Shot Subject Style Composition

🔗 Source: arXiv

SubZero: Composing Subject, Style, and Action via Zero-Shot Personalization

🚀 Technical Novelty

• Mechanism: Disentangled Stochastic Optimal Controller for iterative latent modulation combined with Orthogonal Temporal Aggregation (OTA) in cross-attention blocks to fuse text, subject, and style features without fine-tuning.
• Nuance: Replaces rigid ControlNet pipelines and per-concept adapter training with zero-order latent optimization and orthogonal feature blending, enabling flexible action prompting and single-reference generalization while strictly decoupling content/style leakage.

💡 Yield

• Establishes new state-of-the-art on face/object-style composition benchmarks (e.g., +4–6% average similarity over RB-Modulation/IP-Adapter) and achieves 64–75% human preference scores while maintaining strict subject/style fidelity without helper prompts.

⚠️ Limitations

• Iterative latent optimization during inference increases computational overhead compared to single-pass methods, potentially constraining ultra-low-latency real-time deployment; performance remains dependent on the quality of pre-trained subject/style projectors.