Figure 3. Overall framework. We extend the visual stream into a frozen Prior Noise stream that preserves the pretrained diffusion prior
and a Control Noise stream that adaptively modulates noise based on the Fidelity Weight to restore fine details, forming a Noise-hybrid
Visual Stream. Leveraging the Fidelity Weight to adjust adversarial signals adaptively, Fidelity-aware Adversarial Training enhances
controllability and achieves one-step inference.