HomeAI News
How to train ChatGPT to write Midjourney prompts? Four stages of AI mapping

How to train ChatGPT to write Midjourney prompts? Four stages of AI mapping

April 18th, 2023
View OriginalTranslated by Google

Recently, I started to try to let Midjourney design the main picture of the blog post, and make some situational pictures and cover pictures in the course briefing. After proficient operation and prompts, it saves time compared to looking for a gallery, and it is more directly suitable for my needs. special picture. In the process of using it like this, although I didn't want to seriously study Midjourney's prompts at the beginning, I unknowingly accumulated a spell note of my own AI drawing tool, including what parameters can be used, used What kind of keyword structure to obtain results that are more in line with your needs. However, when Midjourney is used to generate pictures (article pictures, presentation pictures) every day, I hope that the process of "designing prompts" can be faster, and I think of using ChatGPT to quickly generate them . After trying for a while, I found that ChatGPT can be trained to automatically write the Midjourney prompts I need, and:

  • Many times the trained ChatGPT can describe the picture scene and artistic style better than me.
  • The trained ChatGPT also uses parameters of professional photography and 3D technology better than me.
  • The trained ChatGPT will also download the parameters that Midjourney may need.
  • The well-trained ChatGPT allows me to generate the above-mentioned picture themes and scenes, automatically select the appropriate art style, and automatically download the required technical parameters or Midjourney parameters as long as I think of a "concept".

In today's article, I will share with you my own process of using ChatGPT to speed up the process of learning and applying Midjourney prompts.

Phase 1: Let ChatGPT translate Midjourney Prompt: start simple

First, my major premise is:

Try not to complicate things if they can be solved simply.

The meaning here is that although some of Midjourney's powerful pictures are completed with powerful prompts (it looks complicated, has various keywords, and various technical parameters). However, there are many times when simple prompts can produce good results. For example, I simply issued the following command: "Sunrise at Alishan, Taiwan. --ar 16:9 --v 5", even if I did not use any photography terminology, it can actually produce usable (Alishan) sunrise photos .

Therefore, in the first stage, we can use ChatGPT to "translate" our prompt concept without thinking too complicated. I will first create a dedicated ChatGPT dialogue channel, and then all questions related to Midjourney prompts will be asked in this channel, and slowly build up the ChatGPT drawing prompt assistant I need.

The "translation" here is to translate our Chinese concepts into English drawing prompts. If this can produce the photos you are satisfied with and need, then there is no need to write Midjourney's prompts too complicated.

Phase 2: Let ChatGPT rewrite Midjourney Prompt: From concept idea to image language

But sometimes, more professional prompts can indeed produce the special pictures you need, or make the pictures look better. For example, I provide such prompts: "Sunrise at Alishan, Taiwan, captured in a stunning photograph, featuring majestic mountains and lush forests, birds-eye view perspective, inspired by professional travel photography, breathing beauty, vivid colors of morning sky, peaceful atmosphere, realistic style, high dynamic range (HDR), long exposure, optimal camera settings for perfect sunrise capture, golden hour lighting, --ar 16:9 --v 5」 In such midjourney prompts, in addition to the sunrise in Alishan, Taiwan, I also emphasized the style of the photo (professional travel photography, breathtaking beauty), the angle of shooting (bird's eye view, distant view), desired color, and photographic photos Terminology (HDR effects) and more. The photo effect produced at this time has indeed been upgraded to another level.

But here comes the question, if I am not a professional photographer or draftsman, how do I know how to effectively describe various scenes, details, and techniques of a picture? At this time, you can use ChatGPT to "translate".

Please ChatGPT to "translate" my ideas into better AI drawing prompts.

For example, I would ask: "I want to use AI drawing tools to create a cartoon-style version of Hayao Miyazaki's personal photo, please suggest which English keywords I can add to the prompt? " In this way, I can get a lot of technical terms that I didn't think of and didn't know before , and I can use them in Midjourney prompts. Or another kind of "translation" is to help me transform a simple concept into the scene, composition, and details needed for a picture . I can ask:

  • "Please design an English prompt for "Modern port, real photographic photos" based on the structure and logic of all AI drawing prompts you know, and set appropriate photography technical parameters. "
  • "Please design a "quiet" English prompt based on the structure and logic of all the AI drawing prompts you know. "

Many times, I'm actually stuck when I don't have any ideas for drawing pictures, only a desired theme (for example: modern port, real photography photos), or a desired concept (for example: quiet), at this time, I can use ChatGPT to help I translated into a complete drawing prompt. Although ChatGPT doesn’t actually know Midjourney in its database before 2021, as long as I ask him “Please use all the AI drawing prompt structure and logic you know”, then he will really seriously describe it from various angles. Elements of an image , such as the harbor photo below, are translated by ChatGPT:

  • Landscape detail of the port.
  • The key to taking realistic photos (eg during golden hour).
  • It also helped me provide the aperture, ISO and other details needed for photography.

I just said "modern port, real photography photos, photography technical parameters", and ChatGPT provided me with a more complete and detailed description of Midjourney's needs through translation.

Phase 3: Let ChatGPT imitate Midjourney Prompt: replace expert grammar with my needs

The previous "translation" is more like ChatGPT's own imagination. He can describe our concept as a picture language, but is it suitable for Midjourney? In fact, ChatGPT does not know , so sometimes the prompts produced in the second stage will not produce good results when used on Midjourney. In Midjourney, there are indeed some grammar, structure, and logic that can produce more effective pictures. If you want to let ChatGPT who does not know Midjourney learn, the best way is to let him "imitate" . For example, one time, I saw a great fantasy-style image in Midjourney's exploration gallery , so I first copied the cue from that image. (I suggest that you can often go to Midjourney's exploration gallery to find the high-quality pictures you need and collect their prompts for use)

Then, I asked ChatGPT to "imitate" the prompt of the picture and design a Midjourney command of the theme I wanted. I gave the command like this: "Please help me write a About XXXXX's AI drawing tool English prompt: " The key here is to remind ChatGPT to "think according to the logic of the example":

"Think according to the logic of examples", "Think according to the logic of examples", "Think according to the logic of examples"! It is very important, so say it three times, if it is another prompt, it may become just a replacement of keywords.

In this way, ChatGPT will start to understand the design process behind the example, and use the same process to design different content. (Similar to the thinking chain prompt method mentioned in my teaching: " ChatGPT Prompt Manual: General Three-tier Structure and 9 Skills to Improve AI Productivity ")

It can be seen that ChatGPT will indeed re-correct the content I need according to the "logic" of the example prompt, and "not just replace the keywords". For example, the original example scene is the ocean, and ChatGPT automatically converts the scene according to my theme Changed to City and Sky (which I didn't order), and reworked the entire description. So the following picture is produced.

We can collect many "current best examples" of the type of prompts we need . For example, I have collected prompts for posters, photos, abstract concepts, characters, food, etc. Whichever type is needed, I asked ChatGPT to imitate according to the best example, and then generate hints for new topics I need.

Stage 4: Let ChatGPT understand and create Midjourney prompts: Disassemble the instruction logic

In the third stage above, actually, for me who was not very good at designing Midjourney prompts at the beginning, or for me who did not major in art, drawing, or photography, I could ask ChatGPT to design a professional Midjourney in a decent way Prompts, and produce the picture effect I need. But I also want to save time and learn Midjourney prompts more effectively, so I go further:

I started to "feed" ChatGPT some reminders of Midjourney photos that I like very much, but this time I asked him to "disassemble the structure and logic inside".

I would ask like this: "Please help me analyze the structure of the following AI drawing prompts, and present them in a table: Paste Midjourney prompts example"

Doing so has two major benefits:

  • For myself, I began to understand that the masters of Midjourney used such a structure and logic to design prompts ! I also began to understand some key techniques and important parameters of prompts better.
  • For ChatGPT, the ChatGPT in this dedicated AI drawing channel has begun to understand how to use these structures and logics to design Midjourney prompts !

It is not necessary to feed ChatGPT many examples of prompts at one time. When I occasionally see great Midjourney pictures and prompts, I feed him to ChatGPT to let him produce some fragments of learning. What's more important is: occasionally update the prompts that you think are good recently. After all, the memory of ChatGPT is not permanent . As long as you keep updating, you can have a better design logic when you use it next time.

Then, when I want to generate a Midjourney prompt for a picture, I will ask ChatGPT for help like this:

Please design a "XXXXX" AI drawing tool prompt based on the structural logic of all the prompts you remember.

At this time, I don't need to say too much, and I don't even need to say that I need to use English. ChatGPT will automatically produce useful Midjourney prompts based on the structure and logic of the various prompts memorized in the previous clips. We can see that:

  • I only gave the word "quiet", but ChatGPT automatically depicted a quiet scene detail .
    • However, this is what the aforementioned "stage two" has the opportunity to do! But the following are tips that can only be achieved after learning in stages three and four.
  • And ChatGPT will take the initiative to add descriptions of picture art style and technical skills (these are very important for Midjourney), for example, this picture is an illustration style.
  • And ChatGPT's description of the picture will have more details , such as lens focus, color description and other details.
  • Even Midjourney's exclusive parameters will be added , such as v5 engine, image ratio, quality enhancement, style enhancement and so on.

What if, after studying like this, we ask ChatGPT to design a Midjourney prompt of "Modern Port, Real Photography Photos"? At this time, the whole prompt will become more professional, with more and more accurate descriptions of photography technical language, and more professional descriptions of light and shadow. Of course, there are also special parameters of Midjourney, and finally a picture that I am more satisfied with is produced. Port photo map. If you compare it with the pictures generated in the previous stage 2, the port map generated in stage 2 is actually not bad , but the point is that the photos produced in stage 4 are more like the "real photographic photos" I want. This is after ChatGPT learning result .

Finally, I will provide a few application cases of stage 4 for your reference. At this time, I can give some abstract concepts more directly and let ChatGPT help me design useful Midjourney prompts. For example, if I say I want the cover of the presentation, it will automatically add the midjourney image ratio parameter --ar 16:9 for me.

Don't forget that ChatGPT is a dialogue mechanism. If you find that the effect is not good after testing the Midjourney prompts, just use the dialogue in ChatGPT to ask him to correct it.


What I want to say is that the aforementioned four stages of training ChatGPT to help me write Midjourney prompts do not mean that stage four will be better than stage three! So sometimes a simple sentence, ChatGPT translated into English Midjourney prompts, is the best way (stage 1)! However, we can also see from the different training methods of stage 2, stage 3, and stage 4 that using ChatGPT to write Midjourney prompts can also have these flexible methods, so that we can design more creative pictures that meet our needs.

  • Phase 1: Let ChatGPT translate Midjourney Prompt: start simple
  • Phase 2: Let ChatGPT rewrite Midjourney Prompt: From concept idea to image language
  • Phase 3: Let ChatGPT imitate Midjourney Prompt: replace expert grammar with my needs
  • Stage 4: Let ChatGPT understand and create Midjourney prompts: Disassemble the structural logic

Moreover, in this learning process, in addition to generating the pictures I need more quickly every day, I also gradually learned the key structure, logic, and parameters of Midjourney prompts (and organized them into notes). Sometimes, I used Stage 1 to write them myself. Prompt words, but can achieve more desired effect. And want to know which prompts can be set in Midjourney? What are the different meanings of different parameters? The best way is to refer to this official teaching document of Midjourney first , and you will learn a lot . Basically, the structure of the Midjourney prompt is very similar to the structure of the ChatGPT prompt , both of which are: purpose priority, input data, and set output in three stages:

  • Purpose first: picture subject, scene, character description.
  • Input data: Set the picture, details of the photo, such as composition method, photography or painting technique used, artistic style, time period, etc. (especially the artistic style has a great influence)
  • Set the output: Use various parameters to set the picture aspect ratio, quality, artistry and other details.

Also for your reference.

Reprinted from Esor HuangView Original


no dataCoffee time! Feel free to comment