論文研究關鍵字

首頁Home › 人工智慧與新興科技 AI & Emerging Tech

NO.804

文字生成圖像

Text-to-Image

1這是什麼What is it?

文字生成圖像(text-to-image)是依據自然語言描述產生圖像的任務,典型系統由文字編碼器(如 CLIP,2021)與圖像生成器組成,前者把語意映射到共同的嵌入空間,後者據此逐步生成。Ramesh 等人的 DALL·E(2021)與 Imagen、Stable Diffusion 為代表;其效果受提示詞的語法、風格詞與負面提示影響,對文字拼寫與精確空間關係的控制仍有限。

Text-to-image is the task of producing images from natural-language descriptions, with typical systems composed of a text encoder (such as CLIP, 2021), which maps semantics into a shared embedding space, and an image generator that generates accordingly step by step. Ramesh et al.'s DALL·E (2021), Imagen and Stable Diffusion are representative, its results are affected by prompt syntax, style terms and negative prompts, and control over text spelling and precise spatial relations remains limited.

2常被用在哪種研究問題Typical research questions

探索型Exploratory影響/關聯型Effect / association

常見問法:使用者如何撰寫提示詞並評估結果?

示範寫法Example

文字生成圖像的結果取決於輸入。本研究系統操弄提示詞特徵,並分析生成圖像在內容與風格上的差異。

Text-to-image results depend on the input. This study systematically varies prompt features and analyzes differences in content and style of the generated images.

撰寫內容,非真實研究結果,請與指導教授討論後改寫。

Written as a model example, not a real finding. Please revise with your advisor.

3怎麼測它How to measure it

有聲思考、提示詞日誌分析

Measurement notes are written in Chinese. Switch to 中/EN to read them.

4代表文獻Key references

英文論文(2000 年後)English paper (post-2000)

Liu, V., & Chilton, L. B. (2022). Design guidelines for prompt engineering text-to-image generative models. CHI '22, ACM.

中文論文(2000 年後)

中文論文為預填檢索連結,請自行挑選一篇 2000 年後的期刊論文並核對作者、年份、出處後再引用。

英文引用來自模型整理,引用前請務必核對並閱讀原文。English citations are compiled by a model. Verify and read the original before citing.

‹ AI繪圖AI Image Generation提示詞工程 ›Prompt Engineering