Possibly incorrect image normalization step in examples/dreambooth/train_dreambooth.py

[john-sungjin]

Describe the bug

I was studying the train_dreambooth.py script (link) and noticed that the DreamBoothDataset performs an image normalization step as one of the transforms (line 260), setting mean to 0.5 and standard deviation to 0.5:

        self.image_transforms = transforms.Compose(
            [
                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]), # <--- this line
            ]
        )

However, this seems to differ from the original Stable Diffusion training implementation, which instead rescales image pixel values from [0, 255] to [0, 1] (link, line 369):

        example["image"] = (image/127.5 - 1.0).astype(np.float32)
        example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32)

Is there a reason for the discrepancy? Was wondering if this was intended or a bug - it seems like changing the normalization procedure from the original model would impact fine-tuning performance. Thanks for the help!

Reproduction

No response

Logs

No response

System Info

Latest version

[john-sungjin]

Realized I was doing the math wrong; the normalizations are definitely the same! My mistake.

[patrickvonplaten]

No worries :-)