+input? -input? What comparator? If it's LM393, et al., what pullup?
It looks like, for the few microseconds during which the LED is recharging, the comparator is linearly coupled from output back to input through a low-gain path; this doesn't matter because the comparator has more gain. It also has plenty of phase shift, so NFB tends to create pretty PWM waveforms rather than smooth operation.
The LED isn't really forced to charge, it's just acting as a series coupling capacitor in a momentarily unstable loop.
If R3 were smaller, you'd couple less, but you'd also have less signal, which means you need more gain anyway, same problem.
Two things you could try: positive feedback applied via capacitor, which will increase the minimum pulse width from ~100ns to something more useful; else, perhaps an op-amp, wired so as to force the LED to a constant voltage. The error voltage generated in the loop corresponds to the LED current blips. The advantage would be, you can force the LED to some particular condition, such as constant terminal voltage, which may have slightly different behavior (response, time constant, whatever).