[linusekenstam]

We've built a few adapters and the layer between us and the model(s) makes it easy to swap out or support multiple.

Currently we are running 100% on OpenAI, GPT3 & Dalle2, but we have tried some (unofficial MJ solutions) and our own trained SD models. We are working on a feature we call "magic photos" that essentially allows you to upload a set of images of a human, train your own model, and use that in as many stories as you want, you can then create unlimited variations of that character in different situations.

Here are some early versions of that. My daughter as a scientist, in her sci-fi lab. She is created on SD using around 12 images for training, hosting the model ~5gb directly on S3. Then we just use GPU provider pay-per-use to generate the images based on prompts. https://twitter.com/LinusEkenstam/status/1617675627695583232...