Can we train AI on our brand?

AI models like DALL-E, Midjourney, and Stable Diffusion can ‘imagine’ almost any concept you can think of that’s in the public domain. They’ve trained on more images than any human artist could encounter in a lifetime, and they know how to draw all of them. However there are plenty of concepts that they don’t yet understand, and often these more niche concepts are the most important for business use cases. For example unless you’re Coca Cola, the popular AI models are unlikely to have seen many images of your brand or product in the training set. If you’re inventing something new it’ll definitely not be present in the models parameters. However there is a technique for ‘training’ the model to understand a specific concept you dictate. It’s called Textual Inversion, and so far it only works only on Stable Diffusion, the open source alternative to DALL-E. The process is that you upload 3 - 5 images of a concept – an object or style – and run the training script. That finds a space in the existing model that corresponds to the combination of features your sample images have, and then adds a label there, a token, so you’re able to reuse that pointer next time.

Say for example you wanted to add the concept of a specific product you were about to launch called the GoolyBib. You could upload 3 - 5 images and train the model for 1-4 hours on a GPU, a specialised graphics card suited to running ML models, and then that’s saved to a concepts library. Once you import the library you can reference that concept again using the token you designated, for example <goolybib>. Now in prompts you can say “<goolybib> photoshoot in high detail”, or “photo in street style of <goolybib>”, or “elon musk buying a <goolybib>”, whatever images you need to create. You can also train styles, so for example rather than training on your product, you could train on a consistent brand style, and use that in your images to make sure they’re recognizable and easier to get approval to use. This solves a key issue with AI models that have precluded them from being used by businesses in great numbers: you need consistency in order to make the images useful.