acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

While image generation and editing have progressed significantly, existing models still struggle with ambiguous intent, complex reasoning, and out-of-distribution (OOD) tasks due to static knowledge gaps and reasoning limitations. To overcome these bottlenecks, we propose RS-Gen, a plug-and-play, training-free, multi-stage agentic framework for image generation and editing. By introducing a closed-loop "question-posing and problem-solving” mechanism, RS-Gen autonomously plans retrieval actions to bridge information gaps and execute deep logical reasoning. Experiments show that RS-Gen yields absolute gains of 0.31 (on a 0–1 scale) and 19.7 over Qwen-Image and Qwen-Image-Edit-2511 on WISE_Verified and RISEBench, respectively, achieving new SOTA results among open-source models and considerably pushing the capability boundaries of foundational backbones.

Primary area:
generative models
Keywords:
Image Generation, Image Editing, AI Agent, Retrieval-Augmented Generation, Visual Reasoning, Training-free Framework

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Q5MvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

A training-free agent framework is proposed for image generation and editing, which reframes image generation and editing into a multi-agent collaborative self-correcting closed loop. The proposed agent framework is plug-and-play, compatible for most image generation and editing models. Experiments on different benchmarks demonstrate the effectiveness of the proposed framework, which significantly alleviates problems such as visual hallucination, contextual co-reference ambiguity, static knowledge cut-offs in image generation.

Critical strengths
  1. This paper gives a detailed discussion of related work and compares the difference between the existing work and the proposed one. In this paper, related works are discussed by grouping into different categories, image generation and editing models, unified multimodal understanding and generation models, and image agent systems. For works in each category, the paper analyzes their drawbacks or challenges. By doing this, the advantages of the proposed framework are more clear.

  2. The proposed agent framework achieves impressive performance on image generation and editing compared to baseline models. The proposed agent framework yields an absolute leap of 0.31 over the baseline Qwen-Image on WISE Verified benchmark, and surpassing Qwen-Image-Edit-2511 and Mind-Brush by 19.7 and 14.4 points on RISEBench. These improvements illustrate the effectiveness of the proposed framework.

Critical weaknesses

This work is more like an engineering practice rather than a research or scientific work. The main contribution of this paper is the agent framework, which includes some existing models, e.g., generation models, editing models, VLMs, LLMs, etc. The generation and editing improvement mainly comes from the combination of various existing models' ability, instead of designing or developing a new model, or proposing a new theory.

Questions for authors
  1. What is the main technical or scientific novelty of this work?
  2. In Equation 1, will the input image set I_{in, 1} be empty for the first turn of an image generation task?
  3. What is the inference time of the proposed framework for image generation and editing? and is this inference time more than the one of other models?
Reviewer_z1qGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RS-Gen, a training-free, multi-stage agentic framework for image generation and editing. It consists of four specialized agents for image routing, intent analysis, reasoning and search, and iterative image generation. By combining external knowledge retrieval, logical reasoning, and generation refinement, RS-Gen aims to improve the performance of existing image models on complex tasks. The authors report substantial improvements on WISE Verified and RISE-Bench, along with ablation studies across different components and model backbones.

Critical strengths
  1. Practicality and flexibility. RS-Gen provides a training-free, plug-and-play framework that can be applied to different image generation and editing models without additional fine-tuning.
  2. Promising experimental results. The framework achieves substantial improvements on WISE Verified and RISEBench, and the cross-backbone experiments suggest potential generalizability.
Critical weaknesses

1. Limited methodological novelty and insufficient component-level validation.

The proposed framework shares considerable similarities with existing works such as Mind-Brush, particularly in intent analysis, question decomposition, retrieval-augmented reasoning, and prompt refinement. Although RS-Gen introduces additional mechanisms such as defensive questioning, cascaded retrieval, adaptive fallback, and iterative refinement, their distinct methodological contributions remain unclear. Moreover, the current ablation studies mainly evaluate the cumulative effects of adding agents, without isolating the effectiveness of these specific mechanisms. This makes it difficult to assess the actual novelty and contribution of the proposed framework.

2. Insufficient evidence for the necessity of the multi-stage agent architecture.

The reported improvements may largely result from the use of powerful external models (Doubao-Seed and GPT-5.4), retrieval tools, and additional generation attempts, rather than the proposed multi-stage architecture itself. A key missing baseline is a simpler single-agent approach equipped with the same models, tools, and comparable inference budgets. Without such comparisons, it remains unclear whether the multi-stage design provides meaningful advantages over simpler alternatives. The lack of inference cost and efficiency analysis further limits the interpretation of the reported gains.

Questions for authors
  1. Methodological novelty: Could the authors clarify the key methodological differences between RS-Gen and Mind-Brush? Can targeted ablations be provided to demonstrate the individual contributions of defensive questioning, cascaded retrieval, and adaptive fallback?
  2. Baseline comparison: How would RS-Gen compare against a simpler approach where Doubao-Seed or GPT-5.4 directly performs task decomposition, retrieval, and prompt rewriting? In particular, could the authors evaluate a single-agent baseline using the same models, tools, and comparable inference budgets, including generation and refinement attempts?
  3. Validation of claimed capabilities: The experiments mainly focus on knowledge-intensive generation and reasoning-based editing. Could the authors provide direct evidence for other claimed capabilities, particularly multi-turn intention understanding, real-time knowledge grounding, and robustness to retrieval failures?
  4. Reproducibility: The Intent Analysis Agent prompt in Appendix A.3.2 appears to duplicate the Image Router Agent prompt and does not match the formulation in Section 3.2. Could the authors provide the correct prompt and representative end-to-end execution traces, including intermediate prompts, tool calls, and refinement decisions?
  5. Efficiency and cost: Could the authors report the average number of generation/editing attempts, reasoning and search calls, inference latency, and API cost per sample? These statistics would help assess the practical efficiency and computational cost of the reported improvements.
Reviewer_XsRbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RS-Gen, a training-free framework for image generation and editing that coordinates four agents for image routing, intent analysis, reasoning and search, and generation with iterative verification. The agents use Doubao-Seed-2.0-Pro as their controller, with GPT-5.4 providing external reasoning and image verification, and Qwen models performing image generation and editing. The framework combines prompt refinement, external retrieval, and corrective generation without updating model parameters. Experiments report improvements from 0.51 to 0.82 on WISE Verified and from 19.4 to 39.1 on RISEBench, together with component and cross-model ablations.

Critical strengths
  1. A practical framework that improves existing image models without additional training.
    RS-Gen can augment existing generators and editors through image generation models and tool APIs. Experiments across multiple image backbones and two controller models provide useful evidence of its adaptability.

  2. Substantial improvements on knowledge-intensive generation and reasoning-based editing.
    The reported gains over the underlying Qwen models are considerable on both benchmarks. The progressive ablations also show that retrieval and reasoning provide substantial improvements, while generation-time verification and correction offer additional benefits.

  3. Explicit handling of uncertainty and intermediate failures.
    The framework specifies how to defer unsupported interpretations, identify missing information, disambiguate entities before retrieving images, and recover from unsuccessful retrieval. These mechanisms address practical failure cases in tool-assisted visual generation. The appendix provides detailed role-specific prompts and implementation settings.

Critical weaknesses
  1. The experiments do not isolate the benefit of the multi-agent design from the capabilities of its constituent models and tools.
    RS-Gen introduces a strong controller, GPT-5.4 reasoning and verification, external retrieval, and additional generation opportunities. However, it lacks a single-agent baseline using the same models, tools, and inference budget. The results therefore demonstrate that the complete augmented system improves the image backbone, but do not establish how much improvement specifically requires the proposed four-agent decomposition.

  2. Inference expenditure and computational-budget fairness are unclear.
    Although the appendix specifies maximum iteration limits, the paper does not report inference statistics like average search calls, reasoning and verification calls, image-generation attempts, token usage, latency, or monetary cost. In particular, the gain from the generation agent is not compared against repeated generation or candidate selection under an equivalent budget. This makes it difficult to distinguish improved decision-making from increased inference expenditure, and limits the evidence for cost-effective deployment.

  3. Several claimed agent capabilities receive limited direct evaluation.
    Multi-turn co-reference resolution, retrieval fallback, and adaptive error recovery are emphasized, but the reported experiments do not separately measure their success rates or failure modes. Details of the ablated execution paths are also incomplete, for example, how IR outputs containing unresolved knowledge constraints are passed directly to the image model when downstream agents are removed. More explicit execution specifications and targeted analyses are needed to assess these capabilities and reproduce the ablations reliably.

Questions for authors

Please see weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.