川普遇刺照，用一張2100元？ ！ 文章標題、配圖，AI免費給你來一打

機器之能報道

編輯：Sia

莜譜得多。

據說，自媒體使用川普遇刺照片作為配圖，一張付費 2100 元！

隨便下一張圖片給文章配圖的日子，一去不復返咯。

不過，創作者們的麻煩並未到此為止。

寫完一篇公眾號文章，接下來的十幾分鐘最讓人抓耳撓腮：

標題舖位，黃金地段，寸土寸金。除了突出內容亮點，還要確保標題足夠吸引人點開看，這可比寫文章難多了！

啥？酒香不怕巷子深？

在紅海的公眾號圈子裡，這條行不通。

也別灰心！磨合一段時間，我們發現，大模型在解決這兩個日常挑戰上，還挺管用。

-1-

最短時間起一個合格標題

為了方便嘮，咱就以這篇機器人基礎模型新公司 Skild AI 報導

為例。

如果對文章需要突出的亮點瞭如指掌，只是不知道怎麼整出一個吸引人的排列組合，不妨告訴大模型：

1、這是一篇微信公眾號的文章，需要一個吸引人的標題；2、標題要突出數據優勢、融資額度和機器人GPT

這是Claude 3.5 sonnet 的主意，還不賴！

連亮點有哪些也搞不清楚？

那就先讓大模型幫忙找，再讓 TA 將找到的亮點集中體現在標題上。
Claude 3.5 sonnet 歸納總結的能力，很不錯。

連亮點都懶得考慮？那就直接「依葫蘆畫瓢」吧！

有人說，喜歡「本站」那種風格的標題。等等，那是啥味兒？我們還真沒統計過。

讓大模型分析一下：技術性、大 IP 要突出、吸引眼球......

唔，好像是這樣子。

給新報道起了幾個類似標題。還真有內味兒，前三個幾乎可以直接拿來用，除了有點長。

如果更喜歡微博熱搜風格的標題呢？

我們將 7 月 15 日某時點熱搜，作為輸入，讓大模型分析這些熱搜標題的特徵。

接下來也為這篇文章取一個熱搜風格的標題。是不是還可以？至少，咱有了進一步改善的基礎方案（萬事開頭難吶）：

還可以要求更多備選：
有人喜歡專業、嚴肅的財經媒體標題風格，安排上。

我們集合了三年《財新周刊》封面報導標題，讓 Claude 3.5 sonnet 分析。

短、雙關、重大話題、問題導向......，幾個要點都抓到了：

這是大模型為咱們這篇報道起的幾個標題。

內什麼，前三個標題的風格，是不是確有幾分相似？
-2-

模仿生圖，AI 主打一個省

為了避免版權糾紛，越來越多的人開始用 AI 生成文章的封面。

很多時候，比起讓 AI 毫無頭緒地天馬行空，我們更希望 TA 直接模仿一些現成圖片。

一般，我們會先讓大模型看圖說話，給出提示詞，再將這些提示詞丟給圖片產生器。

這是同事的真實體驗：

如果發現提示詞不好使，可以讓大模型（ Claude 3.5 sonnet ）「仔細描述圖片細節和內容」：

然後，將這些

描述丟給圖片產生器：

這張圖片展示了一個獨特而富有對比的場景:

背景：圖片背景是一面灰色的混凝土牆,看起來質感粗糙,有些斑駁和紋理。

座席：圖片中央有一個簡單的黑色長凳或平台,看起來也是由混凝土或某種硬質材料製成。

人物：一個年輕男子坐在長椅的右邊。他穿著深色西裝和高領毛衣,戴著黑框眼鏡。他的姿勢很放鬆,雙手放在腿上。

機器人：在長凳的左側,坐著一個小型白色機器人。這個機器人有一個圓形的頭部,上面有紅色的眼睛狀燈光。它的身體上有一些彩色的裝飾,看起來像是控制面板。

光線：圖片的光線很有特點,從右側打來,在牆上和地上投下了人和機器人的陰影。這種光影效果增添了圖片的戲劇性和深度。

氛圍：整個場景給人一種未來主義和科技感的氛圍。人類和機器人並排而坐,似乎在暗示人機共存的未來。

構圖：圖片構圖簡潔有力,左右對稱,中間留有空白,增強了人與機器人之間的對比和聯繫。

這張圖片透過簡潔而富有寓意的場景設定,巧妙地展現了人類與人工智慧共處的主題,引發了對未來科技發展的思考。

Of course, you can also use Kuaishou The large model of the self-developed Wensheng diagram is available, and it is now available for free for a limited time. The reputation is also good.

However, the large graphable model is not like DALL-E-3 and cannot digest the large amount of description content generated by Claude 3.5 sonnet.

Therefore, when entering the "Creative Overview", it must be shortened and modified:

A simple black bench, with a young man sitting on the right side of the bench. He was wearing a dark suit and turtleneck sweater, and black-rimmed glasses, with his hands on his legs. To the left of the bench, sat a small white robot with a round head with red eye-like lights. . There are some colorful decorations on its body that look like a control panel. The light in the picture is very distinctive, coming from the right, casting shadows of people and robots on the wall and ground.

Then "pad the image", adjust the degree of similarity, and generate the result:

However, for this picture, the effect is very poor.

In addition to the sudden increase in people and machines, if you look closely at one of the characters, you can see that the character has a Lego-like face.
Recently, technology media The verge reported on the first generative biographical film, the protagonist is Brian Eno, the music producer of David Bowie, U2, Coldplay and others.

Every time this movie is played, the movie materials will be recombined according to a special algorithm, so what you see is different every time.

This collage-style cover design caters well to the theme of the movie and also shows the versatility of a biographical figure.

We first let Claude 3.5 sonnet describe the image carefully, and then threw its description to DALL-E-3 to generate the image.

How to say? It feels neat and a bit dull.

We reduced the description of Claude 3.5 sonnet to one sentence, threw it to the large model, and put the picture on it.

This time, the effect of Ketu is obviously better.

We also tried to imitate the creative and unique cover of The Economist.

This time, the visual effect of Tutu (lower right) is better than that of DALL-E-3 (upper right). However, poor spelling ruined the result.

We used DALL-E-3 to generate a similar Styled pictures that express LLM pose a threat to personal privacy and data.

-3-

Text and fingers

are still the weakness of the picture generator

The previous cases have told us:

If the picture must contain words and text, it will be difficult for TA to do it!

You see, Ketu still misspells OpenAI, and DALL-E-3 often does the same.

It is very difficult to correctly display "Siemens" in the picture of DALL-E-3, whether in English or Chinese.

Generate a picture about Siemens Industrial Copilot, the two key words are also wrong.

Although I knew earlier that it involves finger details, the image generator is still prone to problems.

But I never expected that even a picture like a flower picking its nose would be difficult to succeed.

This is the work of DALL-E-3, it is really "amazing".

Throwing it to Ketu, TA was in a difficult position, and he was too embarrassed to put his fingers in front of his nostrils:
As for the gun in his hand, well, it only had six fingers:

This is not surprising.

Image generators usually use diffusion models to reconstruct images from noise and learn patterns that cover more pixels. Naturally, they perform poorly in generating details such as text and fingers.

Of course, this does not mean that text generators must be spelling masters. Although the underlying technologies behind image and text generators are different, they have similar difficulties in spelling and other details.

After all, we still lack basic common sense of the physical world and the language world.

We will bring more AIGC case demonstrations in the future, and everyone is welcome to join the group for communication.

以上是川普遇刺照，用一張2100元？ ！ 文章標題、配圖，AI免費給你來一打的詳細內容。更多資訊請關注PHP中文網其他相關文章！