Here is how to build an AI avatar product

Intro

A few months ago I kept seeing the buzz for generative AI to the point of not being able to resist diving in.

Danny Postma was one who would regularly post updates and actually engage.

I asked him for a clue, and he pointed me to replicate.com which is a service that makes ML easy, they have GPUs running on K8s most likely. The infrastructure is there and they make it easy to deploy.

@levelsio is even using it https://replicate.com/levelsio-org

Write Some Py

You can leverage existing projects on there or make your own. All ML projects in this context rely on a predict.py script (https://github.com/anotherjesse/dream-templates/blob/main/controlnet-1.5-normal/predict.py)

If you want to code your own you can base it off some examples and use chatGPT to help.

After diving deeper into generative AI and starting from here https://replicate.com/blog/dreambooth-api.

Technique

I learned how to customize the models and the results by using special techniques like control net, image to image, etc.

Leaving the AI up to endless constraints would result in inconsistent pictures, but by adding control net you can create structure. This is the first thing I learned when it came to improving results.

Following the guide you will learn that in the API you can pass your own training data, and you will be able to train the model on your users photos.

Examples

Here are some good examples https://github.com/anotherjesse/dream-templates each template has a different approach.

Here is an open source example you can learn from https://github.com/shinework/photoshot, personally how I like to code things, this code is clean.

If you're curious I built this based on all the info I shared here, some blood, sweat, and tears -> https://aiavatarlab.com/

Cheers! Feel free to ask questions in the comments below or DM me on Twitter https://twitter.com/michaelaubry