Let's be honest, we've all seen a regression model get stuck stubbornly predicting the mean, and it's a classic sign of a logic trap, not a code bug. Your model's real problem is that you're handing it a "bad" image along with the very tension value that created that bad result, but you're asking it to magically guess the optimal tension. It has no data for what "optimal" looks like, so it plays it safe and guesses the average to minimize its error. The game-changer here isn't a more complex model; it's reframing the question. Stop asking for the final answer and instead, train your model to predict the one thing it can actually learn from the image: the adjustment needed to get from bad to good.