如何為逼真人物撰寫 AI 照片提示詞

为什么人是最难的学科

观众是面孔和身体的专家。 We notice a slightly wrong hand, an impossible posture or skin that looks like plastic long before we notice a flaw in a building or a landscape. That makes people the subject where vague prompts fail most visibly. The fix is the same as for any photographic prompt, only stricter: describe what a camera would actually record, and leave as little as possible to the model's idea of an average person.

发明人，不要借人

从一个不存在的人开始。描述年齡範圍、體型、頭髮、服裝和活動，並讓細節變得普通。 Never prompt for the likeness of a real private individual, and never generate a real public figure in a situation that could be taken as genuine. Beyond the ethics, many jurisdictions recognise personality rights, which let people control the commercial use of their name, image and likeness. The rules differ by country, and an invented person avoids the question entirely.

描述一项活动，而不是一个姿势

People photographed doing something look natural;人們被告知擺出像庫存圖片一樣的姿勢。 Instead of a woman smiling at the camera, try a woman in her forties reading a paper map at a train station, holding a coffee in her other hand. An activity gives the body a reason for its posture, gives handivity thingan the body a reason for its posture, hold handfaces state the exves exv. task rather than from the prompt.

Posture and weight

真实的身体有重量。 。 image. A single sentence about where the weight sits and what the person is leaning against often does more than several sentences about their appearance.

服装的材质和合身性

Describe clothes the way you would describe any other surface: a creased linen shirt with rolled sleeves, a heavy wool coat slightly too large, worn canvas sneakers. Fabric that fols, strety rinkds the body underneath. Generic words such as stylish or elegant say nothing about material and usually produce clothes that look freshly unpacked.

皮膚，沒有修飾

生成的面孔通常显得光滑均匀，就像经过大量修饰的广告一样。如果你想要自然的效果，用物理术语来说：自然的皮肤纹理、明显的毛孔、眼睛周围的细纹、鼻子上轻微的晒伤、不化妆。描述光线也有帮助，因为柔和的侧光可以温和地揭示纹理，而平坦的正面光则可以消除纹理。避免堆砌“完美”或“无瑕”等术语，这会使结果进一步接近塑料外观。

讓雙手有明確的動作或道具

手是图像生成器众所周知的弱点。最可靠的方法是给他们一些特定的事情做，拿着杯子，放在桌子上，塞进外套口袋里，而不是让他们自由地留在框架中。可见手指越少意味着出错的机会就越少。在使用图像之前，请务必检查双手：数一下手指，检查它们如何抓握物体，并查看它们与手腕的交汇处。

Light on faces

以三点照明传统描述的经典摄影设置，在拍摄对象上使用主光，使用辅助光来柔化阴影，并使用后面的光将拍摄对象与背景分开。你不需要使用行话来使用这个想法。一侧柔和的窗光、阴影一侧的柔和反射以及头部后面明亮的背景用简单的文字描述了相同的布置。相比之下，头顶正午的阳光会在眼睛下方留下深色的阴影，而且很少让人赏心悦目。

Framing and distance

Say how close the camera is. A head-and-shoulders portrait, a figure seen from a few metres away in a street, a small figure in a wide landscape: each one sets the scale at which exps and hands have to beendered. everything.为了获得最自然的效果，具有清晰活动和真实环境的中等框架是最宽容的起点。

自覺規劃多樣性，而非隨機碰運氣

Image generators reflect patterns in their training data, and one study of Midjourney found that even neutral prompts returned unequal results by gender, skin colour and location.如果你想要一個特定的人，請具體描述他們。 If you are making a series, vary age, body shape and appearance deliberately rather than accepting whatever the model returns first, and review the set as a whole before you publish it.

知情同意與生成式合成媒介

Partnership on AI 负责任的合成媒体框架围绕三个理念构建：同意、披露和透明度。对于生成的人来说，同意主要意味着未经许可不得使用任何人的肖像，包括通过参考照片。 If you work from a photograph of a real person, for instance to keep a consistent character, make sure you have that person's clear agreement for the specific use, and never create images that sexualise, humiliate or impersonate hum.

應當了解的標示與披露規定

在歐盟，《人工智慧法案》將深度造假定義為人工智慧生成或操縱的圖像、音訊或視訊內容，類似於現有的人、物件、地點、實體或事件，並且會錯誤地顯得真實。產生此類內容的部署者必須揭露該內容是人為生成或操縱的，對於明顯的藝術、創意、諷刺或虛構作品，揭露更為有限。歐盟委員會表示，這些透明度規則自 2026 年 8 月起適用。在可以被視為真實的背景下，一個現實的虛構人物應該在任何地方都被貼上標籤。

在檔案詮釋資料中標註，而不僅限於說明文

重新分享圖像時，標題可能會被剪掉或遺失。元資料標準的存在是為了在文件中攜帶資訊：IPTC 數位來源類型詞彙表包含使用生成式 AI 創建的媒體的值，並且許多工具會自動寫入此類資訊。編輯和匯出時保持其完整，並在人們可能合理地將圖像誤認為真人照片的地方添加可見的註釋。

LarpGPT 適合的地方

LarpGPT 是一個關於旅行、建築、汽車和生活方式場景的照片提示庫，圍繞著主題、構圖、燈光和鏡頭編寫，並且它的許多生活方式場景都可以使用上述方法進行調整，以包括發明人。您複製提示並將其貼上到您已使用的圖像生成器中。 LarpGPT不產生影像；尊重他人的權利並揭露合成內容由您負責。

延伸閱讀

• Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 3 and 50
• European Commission: AI Act, regulatory framework and application timeline
• Partnership on AI: Responsible Practices for Synthetic Media
• Personality rights (Wikipedia)
• Midjourney (Wikipedia), section on bias in generated images
• IPTC Digital Source Type NewsCodes