Think about it this way instead - every drawing here is an exercise, like a three dimensional puzzle. We know what we're aiming for (based on our reference image), but the real focus of the task is on how we're going to gradually build our construction up, step by step, from simple to complex, in the direction of that goal.

As you step through this process, with every choice you make altering the physical, solid structure that you're creating, you move either closer to that result, or you deviate from it. Regardless of which, you always have to hold to the idea that you're building something real and tangible - and therefore you can't arbitrarily jump back to interacting with your construction as though it's just a drawing to take some shortcuts to bring it back in line with your reference. You simply have to keep going, accepting that some aspects of your drawing won't match the reference - but that it's okay. The goal is ultimately just to have something that feels solid in the end.

The real reason we use reference at all is because students simply don't know what specific things look like. They don't know what plants look like, they don't know what insects look like, they don't know what animals look like. They can identify them by sight, but the information required to reproduce them in a way that is actually realistic or believable, simply isn't there yet. So, we need some source of information to help fuel the choices we make as we work through each of these puzzles.

That doesn't mean we can be lazy in terms of observation - because poor observation will result in an unbelievable, oversimplified construction, which will interfere with the benefits of the puzzle itself (and it's by solving those puzzles over and over that we rewire the way in which our brain perceives the 3D world in which our drawn objects exist). But it does mean that deviation from your reference isn't a huge deal, as long as the effort is being made to observe your reference closely and carefully, and to keep coming back to it to inform every subsequent decision.