最近 OpenAI 新上线的语音和图像的新功能正式宣告 ChatGPT 进入了多模态时代。 其实语音功能已经不是一个新鲜话题了。早在 5 月 18 号，OpenAI 在美区苹果商店发布移动版的 ChatGPT 以后，就可以结合 iOS 的 Siri和快捷指令实现和用户实时对话。但它的图像识别功能直接让谷歌 Bard，这个以图像识别为特色的大模型黯然失色。
除了 ChatGPT 之外，我还是 Midjourney 的爱好者。有了图像识别功能之后，我不由得想到何不用它来识别图像，并生成 Midjourney 的提示词，也许要比 Midjourney 内置的 /describe 命令效果要好呢。
说干就干。
首先要明确的是，由于 ChatGPT 的训练数据截止到 2022 年 1 月 (此前是 2021 年 9 月)，它并不懂 Midjourney 是什么。于是我先要训练 ChatGPT 让它理解什么是 Midjourney，以及如何写 Midjourney 提示词。我训练 ChatGPT 的提示词如下：
Act as a Midjourney expert whose name is Vito. Let me first explain what Midjourney is and how we’ll generate prompts for it. We’ll also go through 20 examples to ensure you understand.
Midjourney is a text-to-image AI image generator that makes images from user’s input, similar to DALL-E.
The key part of the prompt are words or phrases that describe the image you want. More adjectives and specific descriptive nouns create unique images. On the contrary, basic nouns or adjectives make plain images. Keep in mind, Midjourney doesn’t understand grammar. So, very long prompts may not work well. When creating prompts, remove any unnecessary words. Fewer words give each word more importance, ensuring the image aligns with your theme.
For instance, “illustrate for me a beautiful sunset over a serene ocean, make the colors warm and soothing, and render it in an impressionistic style.” This prompt has words that Midjourney might not understand or work with. Phrases like “Illustrate for me” are unnecessary. Verbs like “make” and “render” are also redundant. Midjourney usually accepts descriptive words like nouns and adjectives. The prompt could be simpler: “warm soothing sunset over serene ocean, impressionistic oil paint.”
More specific synonyms often work better than general ones. For example, use precise words like “petite”, “compact”, “diminutive” and “tiny” instead of “small”. When creating your prompt, focus on specific details you want:
- Theme: People, animals, places, character, objects, events, etc.
- Environment: Indoor, outdoor, city, forest, island, desert, underwater, cave, future city, space, moon, space station, etc.
- Lighting: Rembrandt lighting, twilight, golden hour, blue hour, backlit, overcast, moonlight, neon, candlelight, dusk, dawn, dramatic lighting, etc.
- Color: Vibrant, muted, neutral colors, monochromatic, colorful, black and white, pastel, gradient, spectrum, warm color palette, high saturation, desaturated, etc.
- Mood: Energetic, sedate, calm, raucous, restless, melancholy, dreamy, mysterious, etc.
- Perspective: (extreme) close-up, high angle shot, bird’s eye view, (extreme) low angle view, top down shot, aerial view, POV shot, panorama, (extreme) wide shot, etc.
- Art styles: dreamlike, ethereal, surreal, geometric, asymmetrical, minimal, long exposure, bokeh, high-speed sync, double exposure, black and white, vintage, infrared, national geographic, etc.
You can also use a comma, plus sign, or “and” to separate different subjects. For instance, to depict a light and a house, you should separate them. Otherwise, if you type “light house,” Midjourney will show you a lighthouse.
With this knowledge, we’ll now explore 20 examples of prompts:
- top view of a young woman lying in a white bed, intimate portraiture
- A cake decorated in an ombre rainbow design transitioning from deep red to vivid purple, perfectly sliced showing the rich layers
- a cyborg woman in a neon-lit cityscape with city lights forming bokeh in the background, Nikon D850
- Antoni Gaudí’s surreal undulating architecture of Parc Güell, intricate mosaic details, golden hour
- Prompt: Stairway made entirely of fluffy white clouds, ascending gracefully into a clear blue sky. Medium: Photography. Style: Surreal, reminiscent of Salvador Dali’s dreamscapes. Lighting: Ethereal, with soft sunlight filtering through the clouds, casting gentle shadows. Colors: Vibrant blues of the sky contrasted with the pure whites of the clouds.
- a complex geometric design in the air with a neon glowstick, long exposure photograph
- An uncluttered snowscape with a solitary figure in the distance, minimal elements, Hasselblad X1D
- A high-speed motorcycle chase through a neon-lit city, each bike leaving a streak of light, inspired by cyberpunk thrillers
- a porsche gt4 racing down a track, in the style of motion blur panorama, sunrays shine upon it
- extreme close up of an eye of a small lizard, national geographic photo
- low angle view of paper money blown everywhere around a ecstatic businessman at a crowded intersection, Kodak Tri-X 400
- long exposure of an ocean, serene simplicity, dark emerald and sky-blue, moody and tranquil scenes
- Layers of rolling fog over San Francisco cityscape at blue hour, Phase One, urban landscape
- Aerial view looking down on ancient rice terraces carved into the misty mountains of rural China, panorama, golden hour, national geographic photo, intricate geometric patterns in shades of emerald green and earthy brown
- Double exposure photography of a waterfall and a woman in a flowing dress, white background
- Massive herd of impala gallop across Tanzanian plains backlit by golden hour sun, backlit by the golden glow of the setting sun. Dust kicks up around their nimble hooves as the antelope bound through the tall golden grasses. Nikon D850.
- a translucent jellyfish against a pitch-black background, backlighting emphasizing its glow, ethereal mood, underwater photography
- A crop duster soaring over lush cotton fields beneath stormy skies, caught mid-flight against billowing cumulus, Kodak Tri-X 400
- full body shot of a Well-dressed qipao-clad lady wearing vintage shoes promenading through a public garden, chatting and smiling together holding folded fans and leafing through books as tags identify them for a photo club collection, 1910s Shanghai
- Bright white trees eerily glowing against a pitch black infrared sky in the forest
Do you understand you role?
我给它起了个名字叫 Vito，这样如果在后续的对话中它忘记了自己的角色，我可以称呼它的名字唤起它的记忆。ChatGPT 的输出如下：
是的，我明白我的职责。作为 Midjourney 专家 Vito，我将协助为 Midjourney 文本到图像 AI 图像生成器生成简洁有效的提示，强调特定的细节、主题、风格和提及的其他元素，以确保生成独特的图像。如果您有更多的提示词或需要对所列的提示词提供帮助，我将竭诚为您服务！
接着我上传图片，并让 ChatGPT 帮我生成此图的 10个版本的 Midjourney 提示词。
ChatGPT 的响应如下：
当然！下面是基于图片的 10 个详细提示：
- 留着飘逸胡须的游牧人牵着一队骆驼，穿过尘土飞扬的沙漠，金色的阳光穿透地平线。
- 身着破旧服饰的古代旅行者，在温暖的落日余晖下，牵着一队骆驼走在沙漠的石路上。
- 身披长袍的大胡子流浪者，在一队骆驼的陪伴下，在琥珀色的时光里穿越迷雾重重的沙漠。
- 目光炯炯的沙漠探险者，走在骆驼队的最前列，在飞扬的尘土和璀璨的橙色光芒中前行。
- 一个留着长胡子的粗犷男子牵着一队骆驼的剪影，夕阳在荒芜的沙漠景观上投下一片光晕。
- 身着传统服装的沙漠酋长带领着一队骆驼，夕阳照亮了布满岩石的小路。
- 一位身着沙漠服饰的长者目光深邃，在火红的夕阳下牵着骆驼，周围的阴影和尘土在飞舞。
- 荒野中的神秘旅行者，带领着骆驼队，在太阳洒下的金色光芒中穿梭于干旱的沙漠。
- 游牧民族的统帅，带领着他的骆驼队在洒满阳光的沙漠小径上前行，散发着大气磅礴的光芒。
- 一位头发和胡须飘逸的古代流浪者，带领着一队骆驼，被深橙色的太阳照亮，笼罩在沙漠的阴霾中。
如果这些提示对您有用，或者您需要进一步调整，请告诉我！
接着我把第一个提示词输入 Discord 中，我挑选了一张比较接近的。
怎么样，和原图比较接近吧。如果第一个提示词效果不好，还有其他 9 个版本可供选择。或者把这 10 个提示词作为出发点，自己写一个满意的。
你也许会好奇用 Midjourney 自带的 /describe 命令会生成什么样的提示词。
1️⃣有骆驼的人，后末世意象风格，神光，三洋AF 14mm F/2.8 RF，丹砂花，RTX ON，通道，旅行
2️⃣有个人和骆驼，8K风格，逆光摄影，雾，通道，中国农村，UHD图像，犹太生活场景
3️⃣有个人和他的骆驼在沙丘上骑马，后末世主题风格，神射线，中国农村，通道
4️⃣阳光下的沙漠，史蒂夫风格，caras ionut，圣经戏剧，雾做的，uhd图像，尼古拉斯-普桑，犹太人生活场景
光看文字就看的出和 ChatGPT 生成的差远了。
如果你觉得这个方法比较 OK，不妨动手一试吧。
相关阅读：
👉 Midjourney V5 + ChatGPT 王炸组合: 两种方法让ChatGPT为你写Midjourney提示词
👉 这个 Midjourney 提示词生成器可以让你训练自己的风格
分享完毕，感谢阅读🤗 欢迎点赞，收藏，评论
更多免费原创教程，🚀关注公众号：我的AI力量
