# Subject-driven Text-to-Image Generation via Apprenticeship Learning

**Authors:** Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, William W. Cohen  
**arXiv:** [2304.00186](https://arxiv.org/abs/2304.00186)  
**Date:** 2023-04-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.00186v4/

## Key points

- Text-to-image generation has seen significant advancements in recent years
- DreamBooth generates highly customized images of a target subject by fine-tuning an "expert model" for a given subject from a few examples
- This process is expensive as it requires learning a new expert model for each subject
- SuTI is a Subject-driven Text-to-Image generator that replaces subject-specific fine-tuning with in-context learning
- SuTI can generate novel renditions of a new subject in different scenes instantly without any subject-specific optimization
- It is powered by apprenticeship learning, where a single apprentice model learns from data generated by millions of subject-specific expert models mined from image clusters on the internet
- SuTI imitates their behavior to generate high-quality and customized images 20x faster than optimization-based state-of-the art (SoTA) methods
- SuTI outperforms existing models like InstructPix2Pix, Textual Inversion, Imagic, Prompt2Prompt, ReImagen and DreamBooth on challenging benchmarks like DreamBench and DreamBenchv2.
- The human evaluation shows that SuTI significantly outperforms these models on the subject and text alignment aspects.
- Overall, SuTI presents an efficient and effective solution to generating highly customized images of new subjects without requiring expensive fine tuning processes for each individual case.

## Summary

The field of text-to-image generation has seen significant advancements in recent years, with models like DreamBooth generating highly customized images of a target subject by fine-tuning an "expert model" for a given subject from a few examples. However, this process is expensive as it requires learning a new expert model for each subject. To address this issue, the authors present SuTI, a Subject-driven Text-to-Image generator that replaces subject-specific fine-tuning with in-context learning. SuTI can generate novel renditions of a new subject in different scenes instantly without any subject-specific optimization. It is powered by apprenticeship learning, where a single apprentice model learns from data generated by millions of subject-specific expert models mined from image clusters on the internet. Each expert model specializes in a different visual subject and SuTI imitates their behavior to generate high-quality and customized images 20x faster than optimization-based state-of-the art (SoTA) methods. In addition to its speed, SuTI outperforms existing models like InstructPix2Pix, Textual Inversion, Imagic, Prompt2Prompt, ReImagen and DreamBooth on challenging benchmarks like DreamBench and DreamBenchv2. The human evaluation shows that SuTI significantly outperforms these models on the subject and text alignment aspects. Overall, SuTI presents an efficient and effective solution to generating highly customized images of new subjects without requiring expensive fine tuning processes for each individual case.