# Custom Diffusion - Text-To-Image의 Multi-Concept Customization

> Clean Markdown view of GeekNews topic #8161. Use the original source for factual precision when an external source URL is present.

## Metadata

- GeekNews HTML: [https://news.hada.io/topic?id=8161](https://news.hada.io/topic?id=8161)
- GeekNews Markdown: [https://news.hada.io/topic/8161.md](https://news.hada.io/topic/8161.md)
- Type: news
- Author: [xguru](https://news.hada.io/@xguru)
- Published: 2023-01-03T09:48:59+09:00
- Updated: 2023-01-03T09:48:59+09:00
- Original source: [github.com/adobe-research](https://github.com/adobe-research/custom-diffusion)
- Points: 8
- Comments: 0

## Topic Body

- 여러장의 새로운 컨셉 이미지를 제공해서 Stable Diffusion 모델을 미세 조정  
- 키/밸류 프로젝션 매트릭스만 미세 조정하기 때문에 빠름(A100 GPU 2개에서 ~6분)  
- 새로운 객체 + 새 아티스틱 스타일, 여러 새로운 개체, 새로운 개체 + 새로운 카테고리 등의 조합이 가능

## Comments



_No public comments on this page._