Make Any Image Model Tell a Multi-Step Visual Story

*InterleaveThinker* is an innovative multi-agent framework designed to grant standard image generators the ability to produce complex, sequential sequences of text and images. While traditional models often struggle with visual over-reliance and accumulating errors during long tasks, this system decouples the process into a *Planner-Generator-Critic* workflow. A *Planner* agent first establishes a global sequence of instructions, while a *Critic* agent evaluates the outputs and provides corrections to ensure high-fidelity results. This architecture is supported by a robust data pipeline, including specialized datasets for supervised fine-tuning and reinforcement learning. Practical applications of this technology span **visual storytelling**, **robotic manipulation**, and **educational guidance**. Research demonstrates that this collaborative approach significantly boosts the reasoning and creative performance of existing open-source models. *Title, Authors, and Institutions* *Title:* InterleaveThinker: Reinforcing Agentic Interleaved Generation *Authors:* Dian Zheng, Harry Lee, Manyuan Zhang, Kaituo Feng, Zoey Guo, Ray Zhang, and Hongsheng Li *Institutions:* The Chinese University of Hong Kong (CUHK MMLab and CUHK IMIXR) and Meituan *What problem the paper was trying to solve* Recent image generators excel at creating and editing single images but inherently lack the architecture to perform "interleaved generation"—the ability to output a coherent, multi-step sequence of interleaved text and images. While Unified Multimodal Models (UMMs) naturally support this, they fail at long-horizon tasks due to two critical issues: "visual over-reliance" (halting the generation prematurely when an intermediate step visually resembles the final goal) and "step-wise error accumulation" (where minor early mistakes compound step-by-step until the final output is ruined). *What are the paper's key novel ideas?* The paper introduces the first multi-agent framework designed to endow any fixed, off-the-shelf image generator with strong interleaved generation capabilities. To completely eradicate visual over-reliance, the framework completely decouples planning from generation, forcing all instructions to be planned upfront before any visual feedback is introduced. Additionally, to overcome the immense computational costs of optimizing long trajectories involving dozens of generations, the authors introduce a novel dual-reward strategy (combining an accuracy reward and a step-wise reward) that achieves trajectory-level alignment through highly efficient single-step reinforcement learning. *What is the architecture or method they are using?* InterleaveThinker utilizes a closed-loop "Planner-Generator-Critic" multi-agent pipeline. First, a *Planner* agent analyzes the user's input and translates it into a full, multi-step execution plan of text prompts upfront. Next, a chosen off-the-shelf *Generator* executes the current step's prompt to produce an image. Finally, a *Critic* agent evaluates the generated image to identify any anomalies or deviations from the original instruction; if the image fails, the Critic refines the prompt and instructs the Generator to try again until the step is successfully executed. *Why the paper matters* This research proves that complex, sequential reasoning and generation capabilities can be unlocked for existing single-image models without fundamentally altering their base architecture. By retrofitting these models with InterleaveThinker, the framework dramatically outperforms open-source UMMs and achieves results comparable to proprietary frontier models like Nano Banana and GPT-5 on interleaved benchmarks. Furthermore, it surprisingly and significantly boosts the reasoning capabilities of base models on rigorous image generation and editing benchmarks (such as WISE and RISE). *What are the potential applications* Because it allows models to successfully follow a coherent text-image sequence over a long horizon, InterleaveThinker has critical, real-world applications in visual narratives (such as storytelling or character progression), visual guidance (such as generating complex step-by-step instructional tutorials for art or life skills), and embodied manipulation (such as predicting sequential robotic action states like folding clothes). The description, research summary based on a human template and the video were generated by Google's NotebookLM on 29th June 2026