How we increased user retention in the application by 43% with the help of Mid Journey
I am the creator of the service mind tracker. In simple words, Mind Tracker is a rethought diary of emotions. A year ago we had a release. Keeping a diary of emotions is not at all easy, you will, on average, feel the effect in 2-4 weeks. Every day to come in and note the mood, emotions are very lazy, and sometimes even infuriating. Although this is the first sign that the therapeutic effect is already beginning. But in the beginning, not about the therapy itself, but how we tried to keep the attention of users and what we did.
In terms of business, we went from gamification and promotion. For every 20 ratings within the application, the user receives an individual sphere of emotions (for now, I’ll talk about numbers, more about spheres of emotions below). We had a hypothesis that if the user is even a little rewarded for his work, then he will be delayed. And so it happened. Retention 7 days we have increased by 20%, and most importantly Retention 30 days by 43%. We just wanted to increase the retention rate on the 30th day. Also, our virality began to increase, users began to share their sphere on social networks and with friends.
In a nutshell about Neurospheres
In the Mind Tracker application, a person notes his mood, energy, emotions. Based on this data, a graphic representation of your emotional state is created in the form of a virtual sphere. The color of the sphere is your mood lately. The shape of the sphere is energy, the more energy you had, the more active the sphere. The filling of the sphere is your emotions.
Oh promty my promty
Prompt (from the English prompt – “encourage”) is query, hint, or instruction – those inputs that you type when you communicate with the neural network. The correct prompt, that is, the correct request, determines how relevant the output information will be.
The goal is to generate a nice visual sphere based on the user’s state over the past week.
By default, prompt consists of 3 parts: Head, body, tail.
Head
Here, the simpler the start, the more controllable the outcome. I started tritely with “Sphere, glass sphere”. During the experiment, I found two more parameters: “magic”, “abstract” which set the main tone and this is exactly the visualization that suited me. I wanted the sphere to be a little “mysterious” and interesting to look at.
Body
These additional details are already at the heart of the composition. We have 3 main parameters: Mood, energy, emotions.
Mood
For mood, everything is simple, in our application we use a 7-point scale, each mood is tied to a color. For example, bad – red, good – blue. Therefore, everything was clear with the mood. What was the mood of the user most often over the past week, then the word is added to the prompt. That is, in a bad mood, we simply add “red” to the prompt body.
Energy
Next, energy, with energy, I found the following parameters for myself: “Energy inside sphere, electricity” – when a person has had a lot of energy over the past week. “Low Energy, Calm” – if there was little energy.
Emotions
But with emotions it was the most difficult. Firstly, this is a very abstract entity and it is generally difficult to visualize it. Secondly, MJ tried to generate faces for me every time.
Defects. These are the results that the neuron gave me on “angry”. Like I said, it’s very literal. And of course, this is not what a person wants to see in his diary of emotions.
From “angry” I completely abandoned. Going through dozens of words, he stopped right at the base-base: “Calm, Joy, Sadness”. The principle here is the same as what emotion a person experienced more often, then we insert the word into the body of the prompt.
Tail
For the first time when people start “touching” neural networks, they forget about the tail. The tail is most often responsible for the space around the main subject and image quality. The neural network is too literal. And if you write “super quality, 4k” it will give you a really better picture. You can experiment yourself, make a simple prompt without a tail and then add, for example, “4k, 8k, super quality, ultra details, best quality”. In addition to these prompts, there are still quite specific prompts, but with proper use with them, you can also improve the result. I am not going to immerse you in the details about rendering, 3D editor engines, but in a nutshell, it looks like this. There are renders in 3D editors, for example: octane render, redshift, unreal engine – they are considered more realistic and of high quality. The neural network was trained on images with tags. Therefore, when you add these keywords, it makes them look like. If your image implies 3D, then adding them the result becomes better.
Important: head, body, tail are abstract concepts used so that a person can understand the principle of operation of a neural network. The neural network itself has a different logic: it takes into account the “weight” of the word. The earlier a word is in the prompt, the higher its “weight” and the greater the presence of this word in the final image. But at the same time, for example, in MJ there is a separate mechanic for specifying the weight of a word. It is specified through two colons “::”. For example, if you need an image in red – blue, but there should be more red. You can add the following “red::0.7 blue::0.3”. Deapozone from -1 to 1.
General advice from me
When you pick prompts, try all the synonymous words at the beginning. Do not be afraid to make long prompts, after N iterations you will understand which prompts work better, and you can get rid of some of them and the result will practically not change. There is actually a lot of room for creativity here. Nobody will stop you from using prompts like “electricity fire in deep space amazing” and sometimes you can achieve the desired result with the most absurd prompts. Going through dozens of words in the end you find what you need.
Do not use complex language constructions, forget about declensions, almost never use verbs. Nouns work best.
Don’t forget to put “4k”, “ultra details”, “realistic” and the like in the tail of the prompt. They improve the picture.
Find the perfect balance, the shorter the prompt, the easier it will scale in the future.
Technical implementation
I immediately answer the question why not stable diffusion. At some points, SB does not reach the quality level of MJ or incredibly fine tuning is needed. In MJ, I quickly managed to achieve the desired result, which I could not achieve in SB. But at the same time there was the biggest problem, MJ has no source code or API, and all interaction takes place through discord. We took and wrote a bot that sends requests to discord. Through experiments, I compiled a table on which the bot is guided and “glues” all the necessary words.
Did you know that there are people who are already making money on their promt?) They post an image that can be reached by their promt, and you can buy their promt for money to use for your own purposes. We try to be as open as possible and are happy to share our simple table
Initially, the table was 4 times larger, but by reducing it, the result became better and more controlled.
Color is the frequency of choosing a word for prompt. It turned out to achieve a result in which we get almost 0% of defects and unwanted visual spheres. For example, such as in the example above, with a blue evil ball.
We have a standard subscription for $30 per month. Now we have generated 30,000 spheres. Approximately 500 to 2000 spheres are generated per day. And they are constantly generated in the background, it looks like a huge pipeline. The downside is that we can’t make 100% unique spheres. They are quite unique and really are selected depending on what happened to a person during the week. But for example, we cannot take into account the events that occurred in a person. In the future, we want to train SB already on our spheres, to make a markup. For example, if a person took a vacation and completed Atomic Heart full of euphoria, the sphere will already take into account these moments and display the user’s events.
About therapy
At each stage, we consult with psychiatrists and psychologists. At a minimum, we do not want to harm, as a maximum we want to improve the mental state of a person.
The most important thing is the human interpretation of the sphere. Knowledge is directed deep into the person. All that remains is the human-to-human connection. What do you see? What are your feelings? What worries you? Listen to yourself, what emotions and thoughts come to your mind first?
When a person sees a visualization of their emotions, they also experience emotions, which stimulates additional reflection and introspection. This is especially important when the process becomes systematic and you begin to notice changes. For example, your sphere has become darker than before, new elements and forms have appeared – you understand that something is happening with your inner world.
It is also worth considering that emotions are inherently abstract. Visualizing and interpreting them can help you connect with them in a new way and understand their essence. Ultimately, this makes the process of self-knowledge deeper and more productive, expanding opportunities for personal development and self-improvement. By listening to their feelings and thoughts, everyone will be able to better understand themselves and their inner experiences.
people share
Sometimes it is difficult to explain your condition in words, but such visualization can help. You can ask your loved one to find similarities between you and your field. This is another interesting way of getting to know each other and your emotions. People always find patterns. And you may be surprised at what kind of connections your loved one will find. Or you can ask your story followers about it. Believe me, you will be surprised and get those insights that you didn’t think about before 😉 It’s about how the environment perceives you.
We were very surprised how often people share their areas. Yesterday we found out that some of the users made a special chat for themselves where spheres drop off. They rejoice at the first orb they receive. Because of this, people stay and get into the habit of using the diary of emotions more often.
Experiment with it, maybe you will find even more interesting uses. And they certainly are 🙂 We are constantly trying to develop Mind Tracker. We read all reviews and try to respond to all reviews. And always thank you for your feedback. We continue to make a product for you. More and better to come!
Application Link – https://mindtracker.page.link/install
Our Telegram (here we talk about new features, share research and give away future premium features that we are testing) — https://t.me/mindtracker
Okay, now seriously. Here’s how emotion spheres are actually created:
This is a rather complex and highly technological process. In the beginning, the realm of emotions goes through an incubation period. At this moment, the entire mental state of a person is incubated and flows like juice, saturating the sphere. Initially, the sphere is very small, but as it fills up, you constantly have to change the glass frame. As a result, it grows from 3 to 10 meters. Then the production process is divided, depending on the material of the sphere.
hard spheres – our specialists manually polish and paint each sphere, after which they send it to the make-up artist and photographer.
Plasma Spheres – spheres that have a lot of energy, they are observed exclusively in laboratory conditions, under the supervision of scientists, they must be handled with the utmost care.
organic spheres – such spheres are immediately placed in a metal frame. After that, they are taken out on special trucks to the natural environment and there they arrange a photo session.
And only after all these long, painstaking processes, the sphere gets into the user’s phone. And nothing else.
😉
