近期，AI生成的“二次元年轻人的第一次NTR”主题图片广为流传，引发了广泛关注。本文深入探讨了此类图片的生成原理，重点介绍了Google的nano banana模型在保持角色一致性方面的出色表现。文章详细解析了生成此类图片的提示词，包括场景设定、角色关系、服装细节以及情感氛围的营造，并提供了优化的提示词以供参考。此外，文章还建议在AI Studio等平台进行生成，以获得更优的图像质量。

✨ **模型优势**：Google的nano banana模型在AI图像生成中表现出色，尤其擅长保持角色在不同场景下的一致性，这是生成具有情感冲击力图片的关键。

🎬 **场景与构图**：图片通常设定在快餐店等日常环境中，通过前景人物的孤独与背景情侣的亲密形成对比，利用浅景深模糊背景，突出孤独感和戏剧性。

💡 **提示词的关键要素**：成功的提示词需要详细描述场景、角色（包括cosplay细节或拟人化设计）、情感基调（如忧郁、苦涩、讽刺），并强调参考图对手机显示内容和cosplay设计的影响。

🛠️ **生成平台建议**：文章推荐在AI Studio等专业平台生成，认为其相较于Gemini等通用平台，在生成此类图像时能提供更高的质量和更佳的体验。

这几天「二次元年轻人的第一次NTR」的AI图片传得到处都是，我也跑了一组，效果确实很能唤起难受的感觉……[doge]

主要还是归功于Google的nano banana模型在保持一致性上的优秀表现，出图原理就是把原画喂给nano banana——也就是图里手机屏幕上的那张——然后再以其为基准生成经典的餐厅构图。

我把推友Tz大神的提示词稍微改良了一下，也分享在这里：

<br>A cinematic scene inside a fast food restaurant at night. <br>Foreground: a lonely table with burgers and fries, and a smartphone shown large and sharp on the table, clearly displaying the uploaded anime/game character image. <br>A hand is reaching for food, symbolizing solitude. <br><br>Midground: in the blurred background, a couple is sitting together and kiss. <br>One of them is represented as a cosplayer version of the uploaded character: <br>- If the uploaded character is humanoid, show accurate cosplay with hairstyle, costume, and signature props. <br>- If the uploaded character is non-humanoid (mecha, creature, mascot, etc.), show a gijinka (humanized cosplay interpretation) that carries clear visual cues, costume colors, and props from the reference image (armor pieces, wings, ears, weapon, or iconic accessories). <br>The other person is an ordinary japan human, and they are showing intimate affection (kissing, holding hands, or sharing food). <br><br>Background: large glass windows, blurred neon city lights outside. <br>Mood: melancholic, bittersweet, ironic, cinematic shallow depth of field.<br><br>[reference: the uploaded image defines both the smartphone display and the cosplay design, with visible props emphasized]<br><br>Image size is 585px 1024px<br><br>

* 推荐在AI Studio里生成，体感上比Gemini质量要高。