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

🕒 Published (v1): 2026-06-22 12:09 UTC · Source: Arxiv · link

Why this paper was selected

Multi-stage agentic framework coupling reasoning and search; reusable scaffold pattern for OOD generation

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

RS-Gen is a training-free, plug-and-play meta-harness that wraps any existing image generation or editing model with a four-stage multi-agent pipeline (routing → intent analysis → reasoning+search → generation) to overcome knowledge cutoffs, implicit intent ambiguity, and logical reasoning failures. It introduces a "Questioning-and-Solving" closed loop that pre-resolves knowledge gaps and logical issues before pixel generation begins. On WISE_Verified and RISEBench, it yields +0.313 and +19.70 absolute gains, pushing Qwen-Image variants to open-source SOTA.

Problem

State-of-the-art image generation/editing models (diffusion-based and unified multimodal architectures alike) fail on three axes: (1) coreference and implicit intent in multi-turn interactions; (2) out-of-distribution and long-tail entity knowledge due to static training cutoffs; (3) complex logical/physical reasoning since generation is fundamentally "text-to-pixel" statistical mapping. Existing agentic image frameworks either require expensive supervised fine-tuning (Unify-Agent: 143K tailored trajectories) or use rigid linear pipelines with no self-correction, making them brittle to intermediate failures and impossible to generalize across diverse base models.

Method

RS-Gen decomposes image generation into a strictly ordered four-agent pipeline following "Perception → Analysis → Retrieval+Reasoning → Generation":

1. Image Router Agent resolves multi-turn coreferences and anchors target/reference images using a time-decay + intent-matching hybrid strategy. Critically, a Delayed Resolution principle prohibits the router from filling knowledge or logic gaps with internal priors — it passes unresolved variables downstream. Negative semantic constraints are converted to positive visual descriptions to improve diffusion model compliance.

2. Intent Analysis Agent classifies task complexity to decide between a direct-generation fast path and the full reasoning+search path, abstracting latent knowledge gaps into structured queries.

3. Reasoning & Search Agent operates under the ReAct "Observation-Thought-Action" loop, autonomously invoking external search engines, VQA tools, and logical reasoning tools over multiple steps to produce grounded, logically consistent instructions and reference images.

4. Image Generation Agent executes generation/editing under another ReAct loop, incorporating a "Generate-Review-Correct" closed loop: a Critique tool evaluates alignment and visual quality after each generation, deciding whether to locally modify, iteratively polish, or fully regenerate.

The entire system is training-free and model-agnostic.

Key Contributions

• Training-free, plug-and-play multi-agent harness compatible with arbitrary open-source image models
• Delayed Resolution principle: strict prohibition on hallucinating ungrounded inferences at the routing stage, deferring all OOD/logical decisions to the dedicated reasoning agent
• Dual ReAct loops — one for pre-generation reasoning/retrieval, one for post-generation self-correction
• Negative-to-positive semantic translation protocol to prevent diffusion model negative-guidance failure
• "Questioning-and-Solving" closed-loop architecture as a generalizable paradigm for intelligent generation
• Open-source SOTA on WISE_Verified and RISEBench without any model fine-tuning

Results

• WISE_Verified: +0.313 absolute improvement for Qwen-Image; reaches open-source SOTA
• RISEBench: +19.70 absolute improvement for Qwen-Image-Edit-2511; reaches open-source SOTA
• Qualitative results on: specific real-world entity generation (novel concept cars), logical puzzle-solving (matchstick equation correction), visual pattern inference (question-mark shape deduction), physical state evolution (scene two hours later)

(Full ablation tables and baseline comparisons are in the truncated portion of the paper.)

Limitations

• Multi-agent pipeline adds significant inference latency and API cost per request; not analyzed in the visible text
• Reasoning & Search Agent quality is bounded by external search engine reliability and VQA tool accuracy
• Intent Analysis Agent's binary routing (fast-path vs. full-path) is a single-point-of-failure for complexity classification
• Delayed Resolution boundary conditions (what counts as "external knowledge" vs. safe internal inference) are heuristic and may misclassify edge cases
• No evaluation of failure modes when intermediate retrieval returns incorrect or conflicting information
• Comparison is confined to open-source models; gap vs. commercial products (Seedream, FLUX-2 Max) is acknowledged but not closed

Relevance to Harnesses / Meta-Harnesses

RS-Gen is a textbook meta-harness: it imposes zero training cost on the base model and instead orchestrates a configurable agent topology above it, which is precisely the "harness wraps backbone" architecture pattern. The Delayed Resolution principle is a novel contribution to harness design — it formalizes which decisions a routing/parsing stage may and may not make autonomously, preventing hallucination cascades before they reach downstream executors. The dual ReAct loops (one for pre-generation planning, one for post-generation correction) demonstrate how harnesses can introduce self-correcting feedback at multiple pipeline stages without modifying any component weights. This pattern directly extends the meta-harness design space beyond language tasks into multimodal generation.