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

Abstract: Recent years have witnessed remarkable progress in image generation and editing, particularly regarding instruction following and visual fidelity. However, when handling ambiguous intentions, logical reasoning, and Out-of-Distribution (OOD) knowledge, existing image models often yield sub-optimal results due to a lack of deep reasoning capabilities and real-time external information. Although emerging unified understanding-and-generation models attempt to bridge this gap, they remain constrained by their intrinsic parameter scales and static knowledge gaps. Inspired by agentic paradigms, we propose RS-Gen: a plug-and-play, training-free, multi-stage image agentic framework. RS-Gen innovatively introduces a "Questioning-and-Solving" closed-loop mechanism to accurately identify logical issues and knowledge gaps, autonomously planning actions to bridge information deficits and execute deep logical reasoning. Extensive experiments demonstrate that RS-Gen significantly expands the capability boundaries of foundational image generation and editing models. Specifically, on the WISE Verified and RISEBench benchmarks, RS-Gen yields substantial absolute performance gains of 0.313 for Qwen-Image and 19.70 for Qwen-Image-Edit-2511, respectively, successfully elevating both to the state-of-the-art (SOTA) level among open-source models.

• The paper introduces a plug-and-play, training-free framework that coordinates image routing, intent analysis, ReAct-based reasoning and search, and generate-verify-correct loops to handle complex visual tasks.
• RS-Gen substantially improves benchmark performance, reaching 0.8230 on WISE_Verified and 39.1 on RISEBench, including a rise in logical reasoning from 4.7 to 29.4 over the Qwen-Image-Edit-2511 baseline.
• The framework generalizes across image models, with gains of up to 0.405 on WISE_Verified and 26.4 points on RISEBench, but its effectiveness depends on strong instruction-following models and carefully engineered prompts.

Motivation and problem statement

RS-Gen addresses three persistent failure modes of contemporary image generation and editing models: (1) poor parsing of implicit intentions and coreferences in multi-turn interactions, (2) absence of explicit logical reasoning over physical laws, causal constraints, and visual puzzles, and (3) static knowledge cutoffs leading to hallucination on out-of-distribution (OOD) concepts and long-tail entities. The authors argue that unified understanding-and-generation architectures (e.g., Transfusion, JanusFlow, Bagel) mitigate but do not resolve these issues, since their capabilities remain bounded by parameter scale and training-time data. Existing agentic alternatives are either tightly coupled to specific foundation models with expensive fine-tuning — Unify-Agent requires 143K curated multimodal trajectories — or follow rigid open-loop pipelines, as in Mind-Brush's one-way "think-research-create" flow, which lacks fault tolerance when retrieval fails or generation is flawed.

Architecture

RS-Gen is a plug-and-play, training-free framework that decomposes generation into a "questioning-and-solving" closed loop orchestrated by four sub-agents:

• Image Router Agent: performs multimodal coreference resolution and target-image anchoring using a hybrid time-decay/intent-matching strategy over dialogue history $H_{t-1}$. A "Delayed Resolution" principle strictly forbids the router from filling in external knowledge, state evolution, or logical deductions from internal priors; these are preserved as unresolved variables for downstream agents. It also translates negative constraints into positive visual descriptions to compensate for diffusion models' weak handling of negative semantics.
• Intent Analysis Agent: classifies task complexity into direct generation versus reasoning-and-search routing. For the latter, it emits a structured question set $\mathcal{Q}_t$ via "defensive questioning" (forcing even known non-universal entities through explicit factual interrogation) and "process decoupling" (interrogating only terminal visual features renderable by the diffusion model, bypassing intermediate procedural steps).
• Reasoning & Search Agent: a ReAct-style cognitive hub invoking a tool library $\mathcal{T}$ spanning web search, geographic lookup, VQA, deep reasoning, and image retrieval. A text-then-image cascade anchors ambiguous entities before image search, and multi-round reformulation with visual substitution provides adaptive fallback under API failures or noisy results.
• Image Generation Agent: wraps foundation models as expert tools within an iterative "Generate-Verify-Correct" loop, where an MLLM-based verifier produces structured reports $v_k$ driving correction strategies $\Delta p_k$ until semantic alignment is achieved.

Experimental results

The system uses Qwen-Image and Qwen-Image-Edit-2511 as base generators/editors, Doubao-Seed-2.0-Pro-260215 as the agent driver, and GPT-5.4 as an encapsulated deep-reasoning tool.

On WISE_Verified, RS-Gen achieves 0.8230 overall — a 0.313 absolute gain over the Qwen-Image baseline (0.5100) and 0.195 above the runner-up BAGEL (w/ CoT) at 0.6280, trailing only Nano Banana Pro (0.8760) among commercial models. On RISEBench, RS-Gen reaches 39.1 overall versus 19.4 for the Qwen-Image-Edit-2511 baseline (+19.7), with the Logical reasoning sub-task rising from 4.7 to 29.4; it exceeds Mind-Brush by 14.4 points and trails only GPT-Image-1.5/2 and Nano Banana Pro commercially. On WISE_legacy, RS-Gen scores 0.83 overall, surpassing GPT-4o (0.80) and prior agentic systems including GenEvolve (0.82).

Ablations attribute the largest single contribution to the Reasoning & Search Agent: on WISE_Verified it lifts the score from 0.5850 to 0.7940 (+0.209), and on RISEBench from 26.9 to 35.2 (+8.3). The Image Generation Agent's self-correction loop adds a further +0.029 and +3.9 respectively.

Cross-model generalization is strong: gains of +0.238 (LongCat-Image), +0.405 (Z-Image) on WISE_Verified, and +20.5 (LongCat-Image-Edit) and +26.4 (Seedream-4.0) on RISEBench — more than tripling baseline editing performance. Swapping the driver model to MiMo-v2-Omni yields nearly identical results (0.8120 vs. 0.8230 on WISE_Verified), though RISEBench gains shrink to +13.1, indicating that stronger driver models unlock larger benefits on reasoning-heavy editing tasks.

Limitations

The authors concede that performance depends heavily on meticulous prompt engineering of the multi-agent role constraints and feedback mechanisms, making gains contingent on the instruction-following capability of the underlying MLLM driver; lightweight models may yield limited benefit. Two open directions are proposed: automatic prompt optimization at inference time, and a training-free experience/skill library enabling retrieval-based reuse of successful reasoning trajectories across tasks.

Conclusion

RS-Gen demonstrates that a training-free, model-agnostic agentic decomposition — combining coreference-aware routing, defensive structured questioning, ReAct-driven retrieval with fallback, and self-correcting generation — can elevate open-source foundation models to near-commercial performance on world-knowledge and reasoning-intensive benchmarks. The principal caveat is its reliance on strong instruction-following drivers and hand-crafted prompts, which future work aims to automate.