Input normalization range for DynamiCrafter

[tokki8682]

In DynamiCrafter, I understand that the model expects input images to be normalized to [-1, 1], which is common for diffusion models.
However, in the released code, the [0, 1] → [-1, 1] normalization seems to be commented out:

<enhancer_utils.py>

transform = transforms.Compose([
    transforms.Resize((self.height, self.width)),  # skip centercrop
    # transforms.ToTensor(),
    # transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])

Could you clarify the intended input range for the model, and why the [-1, 1] normalization is commented out in the public code?

[tokki8682]

Additionally,
My understanding is that the base DynamiCrafter model is trained with images normalized to [-1, 1].
In the mvdata.py, I could not find an explicit [0, 1] → [-1, 1] normalization step, but I might be missing something.
Could you please confirm whether the fine-tuning was performed using [0, 1] inputs, or if the normalization is handled elsewhere?

Thank you.

[minsu1206]

I really apologize for my too late response.
I didn't get any notification about this issue...

We commented out that transformation because, for training diffusion enhancer, we just applied ToTensor() for images.
You can find related codes at https://github.com/minsu1206/3DGS_Enhancer_Extra/blob/main/lvdm/data/mvdata.py#L325
Indeed, it would be my implementation mistake. Normalization seems to lead to a better performance.
Fortunately, trained model performs well even though we didn't use normalization here.

Feel free to ask.
Thanks.