A few days back we launched SupaRes which is an image restoration/enhancement platform. What I think is unique about that product is that it offers not only general single-image super-resolution but also a few other enhancements such as automatic white-balance, denoising, tone improvement, or low-light boost.
We have some customers in our regular business [1] who requested something more than "just" image optimization and processing. We've put together a few custom models for them and, after a while, decided to rewrite the whole thing and launch it as a separate product which is now SupaRes.
Most of the neural nets are PyTorch reimplementation of TF models while the others are trained from scratch on synthetic data to boost performance in certain verticals such as real estate or food.
On the backend, we use Node for general apps and microservices which are managed by k8s. Rendering engine is obviously written in Python with Sanic being our framework of choice. We're running on bare-metal colocated in Frankfurt/DE and have two machines with 2 x A40 (48GB) each at our disposal. We are members of NVIDIA Inception program and that helped us secure GPUs at preferential prices. To avoid CUDA OOM every GPU has its dedicated Redis-based task queue where we manage inference requests.
SupaRes is really just a beginning for us in the image restoration space but I'm very happy with what we've built so far. I personally think it looks nice on the frontend and, most importantly, does a really good job enhancing degraded visuals.
Dropping this here as maybe some of you will find the service useful.
[1]: https://optidash.ai