CostaliyA/UniCorn

🦄 UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision

   

🧐 What is UniCorn?

While Unified Multimodal Models (UMMs) excel at comprehension, they often suffer from Conduction Aphasia: the inability to translate internal knowledge into faithful generation.

UniCorn is a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. It partitions a single UMM into three collaborative roles—Proposer, Solver, and Judge—to distill latent understanding into explicit generative signals via self-play.

🌟 Key Features

• —Self-Generated Supervision: No external labels or teacher models required.
• —Cognitive Pattern Reconstruction: Bridges the gap between multimodal "understanding" and "synthesis."
• —UniCycle Benchmark: A new cycle-consistency metric (Text ↔ Image ↔ Text) to validate multimodal coherence.
• —SOTA Performance: Leading results on TIIF (73.8), DPG (86.8), and CompBench (88.5).

🚀 Quick Start

Inference & Best Practices

To optimize generation quality and avoid common pitfalls like blurriness, follow these hyperparameter guidelines:

• —`cfg_text_scale`: Use 4.0–8.0 for balanced prompt following.
• —`cfg_renorm_type`: Use global for general Text-to-Image tasks.
• —`timestep_shift`: Higher values for better layout; lower values for finer details.
• —`num_timesteps`: Standard setting is 50.

📊 Results

UniCorn achieves substantial gains over base models (e.g., +6.5 on OneIG, +5.0 on WISE).

📢 News & Roadmap

• —[x] Jan. 12, 2026: Released model checkpoints.
• —[x] Jan. 07, 2026: Released official Arxiv Report.
• —[ ] To-Do: Release full training and evaluation code.

✍️ Citation

@article{han2026unicorn,
  title={UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision},
  author={Han, Ruiyan and Fang, Zhen and Sun, Xinyu and Ma, Yuchen and Wang, Ziheng and Zeng, Yu and Chen, Zehui and Chen, Lin and Huang, Wenxuan and Xu, Wei-Jie and others},
  journal={arXiv preprint arXiv:2601.03193},
  year={2026}
}

📜 License

This project is licensed under the Apache 2.0 License.