In a new research paper, Hugging Face presents GenRouter, described as the first unified workflow routing framework for agentic image generation. The system addresses a growing problem in text-to-image (T2I) models: while recent agentic workflows add advanced capabilities like external knowledge retrieval and iterative reasoning, they often operate in isolation with fixed, one-size-fits-all topologies. This leads to severe compute mismatch, where simple queries are forced through heavy pipelines.
To solve this, the researchers first formulate GenCanvas, a standardized workflow space that deconstructs the generative process into universal primitives such as search, reason, verify, and sketch, and establishes a scalable library of workflow templates. GenRouter then operates over this unified space, using demand profiling, experience matching, and Pareto filtering to route each prompt to its optimal workflow.
Extensive experiments across diverse benchmarks show that GenRouter achieves superior visual alignment while reducing execution costs by over 95% and latency by 65% compared to heavyweight static pipelines. Moreover, the system continuously self-evolves via accumulated experience, enabling robust zero-shot generalization that boosts performance and halves computational overhead.