Fortunately, or unfortunately, the slowdown isn’t OCR (which is already impressively fast); it’s the DeepL translation that’s killing responsiveness. I know you’re focused on OCR perf, but from my end, that’s not the bottleneck.
Examples (0.4.4)
⬆️ The translation Online with DeepL
⬆️ The translation Offline with Internal Engine
This is one of the many reasons why an internal model is the preferred solution in my use case.