RS-Gen: A Multi-Stage Agentic Framework for Reasoning and Search-Augmented Image Generation

Modern image generation models often struggle with complex, multi-turn instructions, ambiguous prompts, and tasks requiring real-world knowledge or logical reasoning. The authors investigate whether an agentic, multi-stage framework can overcome these bottlenecks—specifically knowledge cutoffs, hallucination, and poor instruction following—without requiring expensive, task-specific training.

RS-Gen treats image generation as a collaborative, multi-stage workflow rather than a single-step "text-to-pixel" mapping. It utilizes four specialized sub-agents:

• Image Router Agent: Resolves coreferences and identifies target images in multi-turn dialogues.
• Intent Analysis Agent: Evaluates task complexity and identifies necessary knowledge or logical steps.
• Reasoning & Search Agent: Employs a ReAct (Reasoning + Acting) paradigm to retrieve external information and perform logical deduction.
• Image Generation Agent: Executes the final generation with a "Generate-Review-Correct" loop to ensure output quality.

This architecture is plug-and-play, meaning it can be integrated with existing open-source models to enhance their performance without retraining the base models.

RS-Gen significantly expands the capabilities of foundational models. On the WISE Verified and RISEBench benchmarks, the framework achieved substantial performance gains, elevating models like Qwen-Image and Qwen-Image-Edit-2511 to state-of-the-art levels among open-source solutions. The framework demonstrates superior accuracy in tasks involving physical evolution, logical puzzles, and the generation of specific, real-world entities that are typically "out-of-distribution" for static models.

This work shifts the paradigm of image generation from static, monolithic models to dynamic, agentic systems. By decoupling reasoning from image synthesis, RS-Gen provides a scalable way to make image models more "intelligent" and factually grounded. Its training-free nature makes it highly accessible for the open-source community, bridging the performance gap between academic models and proprietary commercial systems.