SurroGate: Efficient Text-to-Image Routing via Continuous Generative Profiling in Sparse MoEs

Current text-to-image (T2I) generation routing systems suffer from a severe trade-off between the catastrophic latency of execution-based frameworks and the compromised routing fidelity of lightweight static discriminators. To break this barrier, we introduce SurroGate, a novel T2I router based on the Universal Approximation Theorem (UAT), utilizing a sparse text-based Mixture-of-Experts (MoE) architecture to profile the continuous latent math of candidate image models (e.g., standard Diffusion Transformers (DiT), high-quality latent diffusion (QH), and lightweight distilled diffusion (QL)) with 4 specialized LoRA experts. Component analyses confirm that the profile process successfully resolves the continuous-to-discrete manifold collapse, drastically reducing trajectory approximation error (${\mathcal{L}}_{\text{MSE}}\approx 0.03$) to secure an 86.4% internal routing accuracy. Furthermore, extensive evaluations on the Gecko2K benchmark demonstrate that SurroGate establishes a new Pareto-optimal state-of-the-art: it achieves a 78% Oracle Hit Rate—vastly outperforming the CATImage discriminator (45%)—and maintains strong empirical visual quality ($S=0.65$) with minimum routing cost (${\hat{C}}_{\text{router}}=0.02$). This allows SurroGate to deliver high-fidelity generation while completely bypassing the prohibitive latency and cost bloat of sequential cascades (Frugal-T2I) and concurrent ensembling (T2I-Blender).