Hacker News Reader

Comment by UncleEntity | Hacker News Reader

Text Inversion as a Service?

Yep, that's exactly what it is. I can vouch that it really works. Pricing is $3 to run a fine tune, and then generations are "currently" free and unlimited.

I dig it, because I couldn't be bothered to run a colab notebook, but it is super-fun to play with.

What is the value add compared to other free alternatives?

We're trying to streamline the process and explain it. Running the infra for this is hard. We are running on expensive A100 machines which are not easily available and not part of Google Colab infra. There are lots of hyper-parameters one could toy with, and we're trying to hide those and give optimal results automatically. Obviously this could all just be the start of it, and we might continue to develop on different directions where we see it's most fun :)

This seems possibly almost perfect for me as an API but I can't use it for my service without knowing what the pricing will be for generation after the model is trained.

Like I can't launch with the assumption that will stay free forever unless you really pro.ises that, and can't know if it is feasible to use this instead of setting up everything myself on Google Cloud without knowing that pricing.

$3 sounds pretty good for the training though. How long does it take? Also so you have an option for using a different number of instances to train faster (for more money)?

Hey @ilaksh Let's talk over email. We have a lot of different options and details re the API that we're still considering and trying to improve with our partners

We're relying on Google's Dreambooth paper which differently from Text-Inversion, it does full model fine-tuning. This should help achieve higher fidelity results if you're trying to replicate the distinct features of your subject

Reply on news.ycombinator.com