Logo Jinn AI Generator

Founder's Note

Link: https://t.me/logojinnbot 

Unique logo generator with AI

How it was founded 

It all started when a friend of the founder asked him to generate a specific photo for him. Midjourney did not produce high-quality results. He started researching Stable Diffusion, an open-source technology. For example, one of the tasks was to generate a man in a tracksuit wearing a cowboy hat, with the face of one actor. Midjourney did not work well, or rather, only one element (face, hat, suit) worked well, but it failed to achieve the desired synergy of all image elements.

Still, this task was solved with other tools. There were also attempts to generate other types of images, but it would have been difficult to monetize. After that, the idea of generating logos came to mind. Now, Midjourney and other tools are very bad at generating logos and there have been a lot of experiments on how to improve the generation, but there has been a positive experience with logo generation. We managed to invent a technology that can generate high-quality logos. If no one needs this startup, the technology will be added to the public domain. Or perhaps this technology will be used in the future to further improve the accuracy and quality of images.

Communication with people often helps in projects. They often provide some interesting and new information. In general, the approach to working with artificial intelligence started with lower-level things like math, statistics, designing various low-level models, but now there are attempts to work with higher-level tools that do not require such deep knowledge of low-level ones. 

Main functionality

Currently, the app can generate logos based on a short text description. This is the best tool for inspiration. That is, the bot will not generate the final logo and will definitely not replace designers, it is more for getting ideas and inspiration. It is especially helpful when you have no idea what you want to do, or when even the client does not have it, you can give a link to the bot to generate the desired options, which can then be used by the designer for the final version of the logo.

This service is not meant to replace designers, it is primarily to help designers do their job faster and better. Within 3 hours of publishing the link in Telegram chats, we received more than 100 users who wrote to the bot. So far, more than 1000 generations have been made by more than 250 users.

Technology

We use our own development and leased resources to generate power. When creating the service, we spent a lot of time training models and conducting experiments. The training was mainly performed on an NVIDIA A100-SXM4 video card with 40GB of video memory and 85GB of RAM. The CPU is Intel Xeon 2.20GHz. Logos are generated by a modified Stable Diffusion model. Currently, we are actively working on improving both the generation itself and the product as a whole. Logos are generated on a server with a powerful runpod.io video card. We planned to generate on Amazon servers, but we are still negotiating to obtain permission for at least basic capacities.

Now we face a difficult choice with three models:

1. It sticks to the given promo well, but there are not very good logos – https://drive.google.com/drive/folders/1Uw1pKvJ8E9HTz39y9X4d5pBIBtOL0vp8?usp=share_link 

2. Very poorly adheres to the given prompt, but almost all logos are very good – https://drive.google.com/drive/folders/1EaCZTvF9ORfYPS4ulYUwfWkGuGO1HD2R?usp=share_link 

3. Sticks well to a given promo, but often adds "non-logo" graphics – https://drive.google.com/drive/folders/1QYqSi_6TbOPGJB7G3GzZRSLFyHkql9vl?usp=share_link 

So far, we are using the first option, but we are actively looking for a compromise. We want to find a balance between the models and get a higher-level logo.

Generation 

Its main functionality is to generate logos on request from 1-2 words maximum. The more words, the less accurate the result. You can also just click on the /random button and then random logos will be generated.

Design

To train the network, professional designers were involved, who conducted in-depth analytics and studied dozens of design books, most of which were collections of logos. These designers did a tremendous job - more than 3000 special graphic elements were created - from primitives to more complex shapes, from shapeless elements to elements with perfect shapes. Graphic elements can be divided into 14 main groups, each of which has an additional division into the number of basic elements in one form - from 2 to 12.

Team

Founder: Oleg Kopyl (idea, AI-engineering, telegram bot, DevOps) is a developer and designer who works with well-known global companies. He is the founder of the famous online wills startup iDied.org. 

Marketer: Bohdan Shcherbakov (promotion, analytics, and hypothesis testing).

Marketing

At the stage of testing the idea, only Telegram chats were used, which have already brought more than 250 users. Feedback is currently being collected to improve the service. In the future, a full-fledged marketing strategy will be developed to promote the service.

Competitors

Our main competitors are logo generation services that can be found on Google by searching for "create a logo". That is, these are services that seem to "generate" logos, but in fact they select ready-made template graphic elements from a static list, select ready-made stock shapes, choose a color, add a name, slogan, description, and you're done. 

The main disadvantage of our competitors is that with the same input data, the probability of creating identical logos reaches 90-100%. Then what kind of identification through uniqueness can we talk about? 

Services such as Logomaster, which has a .ai domain, again seem to "generate" options, but in fact it is just a selection from a database of ready-made options with different colors. Or services like Wix, which sometimes gives a 500 error (server error). Or services such as Canva, VistaCreate, Logaster, and others are just relics of the past. Standard, in most cases ugly forms that everyone has already seen and knows, template options, clichés - it's all about the old options and possibilities for generating a logo.

Costs

1. Around $200 for computing power, but we need more ($1000-$2000) because now generations take a long time due to heavy loads and long queues. If the bot is used, we will increase the cost of hardware.

2. For now, we spend $100-$500 on marketing to validate the idea.

What's next?

In the near future, our main plans are to launch on ProductHunt in the near future, test marketing hypotheses, get paid users, increase computing power, improve the quality and variability of logos, and launch full-fledged advertising. 

Soon, logos are likely to be generated in SVG format, which will multiply the bot's capabilities and efficiency. So far, we have received quite a lot of traffic thanks to the reactive viral spread, because people are really interested in this bot. Many people are discussing it, many are already using it. 

We plan to develop both a mobile app and a website, but for now, the Telegram bot is enough to test the hypothesis. In simple words, if we realize that people need this product and are willing to pay for it, then we will improve it.

Link: https://t.me/logojinnbot