A question on Clipping Flow values

[MohitLamba94]

Hello,

I noticed you comment mentioning that Clipping Flow values helped you a lot and default value you set to (-3,3),

rectified-flow-pytorch/rectified_flow_pytorch/rectified_flow.py

Line 159 in 1f59e7e

clip_flow_values: Tuple[float, float] = (-3., 3)

My question is about the range of Clipping values for the Flow.

Let,

gt_flow = X1 - N

where X1 is a image form desired distribution say CelebA dataset normalised in the range [-1,1].
And N is a sample drawn from a Normal distribution. So mostly N will be in the range [-3,3] but theoretically its unbounded so lets say if we collect statistics over several steps most samples will be in the range [-4,4].

So now if we compute max and min of gt_flow which is what network must predict, they are

gt_flow_max = 1 - (-4) = 5
gt_flow_min = -1 - 4 = -5

So just wanted to check do you prefer keeping clip_flow_values = (-3,-3) because your result on Oxford Flower dataset look pretty good, or you keep something more?

If my understanding is correct, this flow clipping happens only during inference and not during training so training is not getting affected.

But even during inference will clipping in the (-3,3) hurt given that predicted flow might need to be (-5,5) or at least (-4,4)?

Thankyou