InterleaveThinker brings interleaved text-image generation to any image model

InterleaveThinker addresses a fundamental architectural limitation of current image generators: their inability to produce interleaved sequences of text and images, which the paper identifies as critical for visual narratives, guidance, and embodied manipulation. Even the most capable open-source Unified Multimodal Models (UMMs) perform poorly on this task. The proposed solution is a two-agent pipeline — a planner agent that organizes the image-text input sequence and issues step-by-step instructions to the image generator, and a critic agent that evaluates each output, identifies samples deviating from the plan, and refines instructions to trigger regeneration.

Because a single interleaved generation trajectory can involve more than 25 generator calls, optimizing the full trajectory end-to-end is computationally impractical.

Training the pipeline involves a format cold-start phase using two supervised fine-tuning datasets (Interleave-Planner-SFT-80k and Interleave-Critic-SFT-112k), followed by reinforcement learning on Interleave-Critic-RL-13k. Because a single interleaved generation trajectory can involve more than 25 generator calls, optimizing the full trajectory end-to-end is computationally impractical. The paper addresses this with accuracy reward and step-wise reward formulations that allow single-step RL — specifically GRPO — to effectively guide the entire trajectory.

Benchmark results show InterleaveThinker achieves performance comparable to Nano Banana and GPT-5 on interleaved generation benchmarks. Notably, the pipeline also produces substantial gains on reasoning-based benchmarks WISE and RISE when applied to 4-step FLUX.2-klein, a result the authors describe as surprising given the pipeline was not explicitly designed for reasoning tasks.

Key facts

1. 01InterleaveThinker is described as the first multi-agent pipeline to add interleaved text-image generation to any existing image generator.
2. 02The pipeline uses two agents: a planner agent to organize and instruct, and a critic agent to evaluate outputs and refine instructions for regeneration.
3. 03Three training datasets are introduced: Interleave-Planner-SFT-80k, Interleave-Critic-SFT-112k, and Interleave-Critic-RL-13k.
4. 04Reinforcement learning uses GRPO with accuracy reward and step-wise reward to make single-step RL practical across long trajectories.
5. 05A single interleaved generation trajectory can involve over 25 generator calls, making full-trajectory optimization computationally impractical.
6. 06On interleaved generation benchmarks, InterleaveThinker achieves performance comparable to Nano Banana and GPT-5.
7. 07Applied to 4-step FLUX.2-klein, the pipeline yields substantial gains on reasoning-based benchmarks WISE and RISE.

Topics