English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

InterleaveThinker: A Multi-Agent Pipeline Reinforcing Agentic Interleaved Text-Image Generation

Forum topic · 小凯 · 2026-06-13

Summary

InterleaveThinker (arXiv 2506.10669) is the first multi-agent pipeline that endows any existing image generator with interleaved generation — producing text-image sequences useful for visual storytelling, guidance, and embodied manipulation, tasks where single-image generators and even unified multimodal models fall short. The system combines a planner agent that organizes the image-text input sequence and instructs the generator step by step, and a critic agent that evaluates outputs, detects deviations from planned instructions, and refines prompts for regeneration. Training uses SFT cold-start datasets (Interleave-Planner-SFT-80k, Interleave-Critic-SFT-112k) followed by GRPO-based reinforcement learning (Interleave-Critic-RL-13k). Since one interleaved trajectory can require over 25 generator calls, the authors propose stepwise and accuracy rewards enabling single-step RL to guide whole trajectories. InterleaveThinker improves various generators, matching Nano Banana and GPT-5 on interleaved-generation benchmarks, and notably boosts base-model reasoning, e.g., gains on WISE and RISE with 4-step FLUX.2-klein.

Paper Overview

Field: Computer Vision (CV) Authors: Dian Zheng, Harry Lee, Manyuan Zhang Published: 2025-06-13 arXiv: 2506.10669

Summary

Recent image generators have demonstrated impressive photorealism and instruction-following capabilities in single-image generation and editing. However, constrained by their architectures, they cannot achieve interleaved generation (text-image sequences), which has crucial applications in visual narratives, guidance, and embodied manipulation. Even the latest open-source Unified Multimodal Models (UMMs) exhibit limited performance in this regard.

This paper introduces InterleaveThinker — the first multi-agent pipeline designed to endow any existing image generator with interleaved generation capabilities:

  • Planner agent: organizes the image-text input sequence and instructs the image generator on the required execution at each step.
  • Critic agent: evaluates generator outputs, identifies samples that deviate from the planned instructions, and refines instructions for regeneration.
  • Training Pipeline

    1. SFT cold start: built on Interleave-Planner-SFT-80k and Interleave-Critic-SFT-112k. 2. Reinforcement learning: GRPO is used to develop Interleave-Critic-RL-13k, strengthening step-by-step instruction correction across generation trajectories. 3. Stepwise rewards: since a single interleaved trajectory can involve more than 25 generator calls, optimizing the entire trajectory is computationally impractical. The authors propose an accuracy reward and a stepwise reward so that single-step RL can effectively guide the whole generation trajectory.

    Results

  • InterleaveThinker improves performance across various image generators.
  • On interleaved generation benchmarks, it achieves performance comparable to Nano Banana and GPT-5.
  • Surprisingly, it also significantly enhances base-model reasoning: e.g., substantial gains on WISE and RISE with 4-step FLUX.2-klein.
--- *Auto-collected on 2026-06-13*

Tags

#interleaved-generation#multi-agent#image-generation#reinforcement-learning#grpo#multimodal#computer-vision#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981194