At the end of 2022, there was an explosion of AI-based products.
Our Sommo team also dived into this rabbit hole. We have created a Dyvo application that allows you to create art drawings and avatars from your photos using AI. Users can generate profile pictures, pet portraits, dating photos, and content for Twitter / TikTok / Facebook / Instagram.
The time from the idea to the first release took less than a month.
The app is built with Bubble for Stable Diffusion v1.5, and training with the Dreambooth.
We also used our no-code development tool Natively to convert the Bubble application to mobile iOS/Android. For Dyvo, having a presence on the App Store and Google Play Market was an important way of user acquisition.
Creating mobile applications with Natively took no more than an hour. In addition, we could customize important native features - social auth, in-app purchases, push notifications, camera, etc.
The challenge of applying Stable diffusion was
(a) to create images similar to the people depicted in the original photos,
(b) avoid artifacts and other bugs associated with AI,
(c) generate images that users will like (the most significant challenge).
To solve this problem, we conducted experiments with training models, prompts, and configs (sampling methods, steps, CFG scale, X/Y plotting, and seeds).
Results
Onboarding
When opening Dyvo for the first time, the user can check the quality of the created AI pictures before signing up. You can compare the original photos and hundreds of images generated by Dyvo for these users. It is handy (see the result in advance) and fun!
You can also learn how Dyvo works and browse hundreds of photos generated and published by Dyvo users (community).
After signing up, users can see (or skip) Onboarding with the main steps to create Dyvo avatars.
Uploading photos
By clicking the “Create avatar” button, the users initiate generating flow: checking out, picking appearance, and uploading their photos.
Sticking to the recommendations when selecting photos is essential to create great avatars. Users should upload selfies with different facial expressions, backgrounds, clothes, angles, lighting, and bigger than 512 x 512 px.
We don’t recommend using the same background/shot, the same emotion, selfies with sunglasses, too much makeup, bikini/swimsuits, or nudes.
Avatar Generation
After uploading photos, the magic starts - the app spins up a machine-learning model individualized just for you. You will get your avatars in 50+ styles in less than an hour. Note that some could be flawed. AI algorithms may generate weird artifacts, inaccuracies, and defects in output images. It’s expected, and the resulting images will never be 100% perfect.
Extras
In addition to generating static pictures, you can:
animate your avatar;
extend it from a small square image to a 4k 2:3 professional photograph;
make your avatar talk - select the avatar, write your text, and we’ll voice your avatar talk with AI video like a real character.
Summary
Within the first month, Dyvo grew by 100k users (android, iOS, and web), and generated thousands of avatars for them. Try creating avatars for yourself or your friend/pet - this is an awe-inspiring experience.
Or try Natively to convert your Bubble projects into mobile apps, release to App Store/Google Play and enable native iOS/Android features.
If you are interested in going deeper into the process of using and configuring Stable Diffusion, there are a lot of headers in the network, for example:
Github Repository with the guides
