Why does the effect deteriorate after increasing the number of Inference Steps for InstaFlow?

#2
by alfinest - opened

According to the paper, the effect should be better and the speed should be slower

InstaFlow is a one-step model after distillation. Multi-step inference is not supported since this is not the 2-Rectified Flow model (it is 2-Rectified Flow+Distill). We provide the 'number of inference steps' option just for comparison with SD.

This means that only the model I trained n-step can perform n-step inference. The effect of n-1 and n+1 is not guaranteed? Thank you for your answer

It is not guaranteed. In my practice, 1-step distillation degrades all the other n(n>1)-step inference.

Thank you for your reply and look forward to open source code

Sign up or log in to comment