Focus on python. It's going to be a MUCH easier barrier to entry. If you see a specific niche you'd like to focus on with C++ then sure - learn that. But I would not recommend starting with C++. I did start there 20+ years ago and wish there was an easier way in back then. My main gripe is there's 100 standard ways to do things in C++ whereas python code tends to be fairly uniform across enterprises / projects / etc.

You don’t think 99% is a super high accuracy?

Your input has to match what the model is expecting. It sounds like this is your problem

There is a fundamental problem. Try having it predict an image after zero examples. This should give you a noisy image from the randomized seed. Then try after 100 examples, 1000, 10,000, etc. My guess is your input images are incorrect so it’s learning to predict all 0s or black. Is the accuracy super high?