InterleaveThinker/Critic-SFT-8B
InterleaveThinker/Critic-SFT-8B is an 8 billion parameter instruction-tuned model developed as part of the InterleaveThinker multi-agent pipeline. It functions as a critic agent, evaluating generator outputs and refining instructions for interleaved text-image sequence generation. This model is specifically trained to enable complex visual narratives, guidance, and embodied manipulation by correcting step-wise instructions, achieving performance comparable to larger models on interleaved generation benchmarks.
Loading preview...
InterleaveThinker/Critic-SFT-8B Overview
InterleaveThinker/Critic-SFT-8B is an 8 billion parameter model designed as a critic agent within the novel InterleaveThinker multi-agent pipeline. This pipeline is the first of its kind to endow existing image generators with interleaved generation capabilities, allowing for dynamic organization of image-text input sequences.
Key Capabilities & Features
- Critic Agent Functionality: Evaluates outputs from image generators, identifies deviations from intended results, and refines instructions to guide the generation process.
- Interleaved Generation: Facilitates complex text-image sequence generation, crucial for applications like visual narratives, guided image creation, embodied manipulation, and long-horizon sub-task annotation.
- Specialized Training: Trained on dedicated datasets including Interleave-Critic-SFT-112k, utilizing GRPO with accuracy and step-wise rewards for instruction correction.
- Performance: Achieves strong results 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, showcasing its adaptability.
Use Cases
This model is particularly well-suited for scenarios requiring precise control and iterative refinement in multi-modal generation, especially where visual narratives or complex, multi-step image creation is involved. It enhances the ability of image generators to follow intricate, evolving instructions.