Midjourney Keyword Harvester
A low-code tool to harvest keywords used by the community
This article is written by Alparslan Mesri and Ugur Ziya Cifci.
With the development of artificial intelligence, art production techniques are also changing. In the process of creating art, new steps and skill sets emerge. The Midjourney platform is one of the new tools created by artificial intelligence for art production.
Midjourney is an independent research lab that features a Discord bot that generates images. It takes human input (keyword prompt) and utilizes a machine learning algorithm trained on massive amounts of picture data to generate unique images. It was used to create the cover image for The Economist in June 2022.
Midjourney users create multiple images by experimenting with different keyword combinations, and they try to get the best image by comparing these outputs.
In order to use these new production techniques efficiently, the importance of choosing keywords is increasing. The process of choosing the keywords correctly and properly in order to increase the quality of the product is called keyword harvesting.
Problem
The problem here is that the choice of keywords is limited by user’s vocabulary, past experience and imagination.
Since there is no guide on keyword harvesting with Midjourney and the machine learning systems used by the Midjourney platform are closed systems (black boxes) that are unpredictable, it is not possible to put forward precise and clear rules for keyword harvesting. In this case, the choice of keywords in creating a digital product is partially up to the inspiration and intuition of the user.
In this article, the low-code keyword harvesting tool we developed as a solution is presented. Thanks to the Tool, user is not only limited to his/her individual imagination, but also benefits from the collective imagination (keywords) of the Midjourney community.
Thanks to this tool, users can see both the frequently used word combinations with the keywords they choose and the rare and unexpected combinations.
For example, let’s say a user wants to generate pictures with different “lighting” techniques, provided the “home” theme remains constant. In this case, the user can see all the combinations used by the community (e.g. sunlight, moonlight, pastel light, etc.) by typing “light” into this tool.
Dataset
The dataset we used was created by the Succinctly AI team, collecting a total of 268k messages from 10 different Discord channels in a 28-day study using Discord bots. You can access the dataset on Kaggle.
The Tool
In this study, we tested the tool by entering “light” in the Search_keyword section and found other word groups used with this keyword according to their frequency of use. Users can use the tool by replacing the “Search_keyword” part with their own keywords.
Case Study
In this study, we took different lighting keywords from the tool. We tested on the “House” theme and different images were generated.
1. House
2. House Cinematic Light
3. House Pastel Light
4. House Liquid Light
We tried “liquid”, one of the most infrequently used keywords.
Most used by the community: moonlight and sunlight.
5. House Moonlight
6. House Sunlight
Conclusion
Sometimes it is appropriate to write the most frequently used word groups, while for more surprising and creative results, it may be better to go with the least used words.
The low code tool we have developed provides important benefits for professional content producers as well as individual users.
When different variations on a theme need to be produced, the content producer can quickly access and “harvest” phrases that have been tried before by the community.
On the Midjourney platform, sometimes works of high aesthetic value can emerge from the most tried word groups, and sometimes from the least tried and unintuitive word groups.
The “liquid light” case is an example of this. In this respect, this tool also offers users the opportunity to be unique at a time when artificial intelligence generates content.
You can access the Kaggle notebook here to try the Tool and see the source code.
