For most image compression tasks, speed isn't the metric that users care about. Since compressing images parallelizes very well, they care about total system cost for a given throughput. So your value proposition is more like "Use 1 GPU box running our code instead of 10 CPU boxes".

Saving 9 boxes isn't worth an engineer spending much time hacking. You should target people with 1000 CPU boxes to switch to 100 GPU boxes (or whatever the ratio is.) Are enough customers like this to justify building a business around it?