GenRouter: Unified Workflow Routing for Agentic Image Generation
Harold Haodong Chen, Zhiyu Hou, Wen-Jie Shu, Weilin Ruan, Yingjie Xu, Litao Guo, Ying-Cong Chen
cs.CV
2026-08-17
HKUST researchers decompose agentic image-generation pipelines into eight primitives and a template library, then route each prompt by demand profile: visual quality matches or beats GEMS while execution cost drops over 95% and latency 65%.
Raw pixel synthesis is largely a solved problem in text-to-image models; the pressure has moved to the request side. "A cat in an astronaut helmet" and "an ICLR 2027 poster with exactly readable title text and three logos arranged left-right" place completely different demands on the pipeline. Academia has answered with a stack of agentic systems: Mind-Brush with search, GEMS with a verification loop, SCOPE with skill orchestration, GenClaw compiling layout to code.
These systems share two diseases. Fragmentation: each builds its own siloed pipeline, and capabilities do not compose. Compute-mismatch: a fixed topology means simple prompts still traverse the full search-reason-verify stack. The paper measures GEMS at 40 dollars and 8.4 hours on GenEval2, against a floor of zero for direct generation. GenRouter's bet: not every request should pay flagship-pipeline prices.
The framework has two layers. GenCanvas, the lower layer, decomposes existing pipelines into eight atomic primitives: Decompose (extract constraints), Search, Reason, Skill (retrieve prompt-engineering instructions), Sketch (generate SVG or HTML layout code), Verify (check against a constraint checklist), Refine (assemble the final prompt), and Rewrite (reformulate after a failed verification). Primitives compose into four template tiers: semantic alignment generates directly or with light rewriting; external grounding searches first; structural reasoning produces layout code first; iterative refinement closes the loop with verification. GEMS, Mind-Brush, and SCOPE all map onto specific templates, and the abstraction ships as an open-source codebase.
GenRouter, the upper layer, dispatches in three stages:
Execution outcomes flow back into memory, so the router improves with use and never needs retraining.
Across five benchmarks with the Z-Image backend plus a Qwen-Image-Edit-augmented variant, the average score is 70.2 versus 68.5 for GEMS, but the cost and latency gaps are an order of magnitude:
| Method (3B-backend average) | Avg score | Cost ($) | Latency (h) |
| GEMS | 68.5 | 37.4 | 14.0 |
| SCOPE | 62.7 | 3.5 | 7.0 |
| Direct generation | 55.1 | 0 | 0 |
| GenRouter | 68.1 | 3.1 | 4.1 |
| GenRouter augmented | 70.2 | 3.0 | 4.4 |
Routing distributions track benchmark character: aesthetics-heavy DPG-Bench sends roughly nine in ten prompts to lightweight templates, while OneIG prompts with spatial and logical constraints use the full template spectrum. In the self-evolution experiment, sequentially accumulating experience from three benchmarks lifts the mixed-set score from 62.5 to 73.5 while cost and latency both fall; freezing experience distilled on WISE and transferring zero-shot to DPG-Bench scores 87.1, above 85.9 for an LLM-as-router, at half the cost. Ablations: removing demand profiling drops performance to the low 70s, clearing the memories sends cost to GEMS levels. All three components matter.
For teams building image-generation products, this is a step from demo to deployable: heavyweight capabilities stay available, but only requests that need them pay. Experience-based routing rather than a trained policy means swapping the base model does not force router retraining, and the recipe transfers naturally to other multi-pipeline agent settings. The open codebase also puts previously siloed pipelines on one comparable ledger, with unified token and latency accounting.
The incremental part is real: on quality alone the gains are a few points; the story is the order-of-magnitude cut in cost and latency.
Cost and latency count only routing and primitive overhead, deliberately excluding the image generator's own inference, which in deployment is often the larger bill. The paper's reason is fairness across generators, but readers must add that cost back. Benchmark scores are cross-dataset calibrated with weights from 1.12 (WISE) to 1.69 (ArtiMuse), so the "70.2 average" is a synthesized number, not an official single-benchmark score. Routing quality depends on an exhaustive cold-start exploration against the target benchmark; a genuinely new distribution restarts the clock, and zero-shot transfer was verified on one path only (WISE to DPG-Bench). The self-evolution comparison is against fixed pipelines, not against a continually retrained model router.