InterleaveThinker/InterleaveThinker-Planner-8B
InterleaveThinker/InterleaveThinker-Planner-8B is an 8 billion parameter model developed by Zheng et al. for agentic interleaved generation. It functions as a planner agent within the InterleaveThinker multi-agent pipeline, designed to endow existing image generators with interleaved generation capabilities. This model excels at organizing image-text input sequences and refining instructions for complex interleaved text-image sequence generation, achieving performance comparable to advanced models on relevant benchmarks.
Loading preview...
InterleaveThinker-Planner-8B Overview
InterleaveThinker-Planner-8B is an 8 billion parameter model that serves as the planner agent within the novel InterleaveThinker multi-agent pipeline. Developed by Zheng et al., this system is designed to enable existing image generators to perform interleaved generation, a process involving complex text-image sequence generation. The model's primary function is to organize image-text input sequences, evaluate generator outputs, identify deviations, and refine instructions through a critic agent.
Key Capabilities
- Agentic Interleaved Generation: Facilitates the creation of visual narratives, guidance, embodied manipulation, and long-horizon sub-task annotation by managing interleaved text and image inputs.
- Instruction Refinement: Works in conjunction with a critic agent to correct step-wise instructions, ensuring coherent and accurate output.
- Performance: Achieves performance comparable to models like Nano Banana and GPT-5 on interleaved generation benchmarks, demonstrating significant gains on reasoning-based tasks (e.g., boosting WISE from 0.47 to 0.73 and RISE from 13.3 to 28.9 on 4-step FLUX.2-klein).
- Transferability: Improves performance across various existing image generators, enhancing their interleaved generation capabilities.
Training Details
The model was trained using dedicated datasets: Interleave-Planner-SFT-80k, Interleave-Critic-SFT-112k, and Interleave-Critic-RL-13k, utilizing GRPO with proposed accuracy and step-wise rewards for instruction correction.