input

[c1ircle]

Why do I input the same image with different resolutions, but the size of the object in the rendering result is different? As far as I know, the preprocess function of the image for different sizes of the image, the processed image is the same, only one is clear and another is blurred.
The input image to

TRELLIS/example.py

Line 17 in eeacb0b

image = Image.open("assets/example_image/T.png")

:
image1（512x512):

image2（2048x2048):

image1 after preprocessing:

image2 after preprocessing:

The output of image1:
https://github.com/user-attachments/assets/6afe39b3-e5e3-49c5-8dbc-b430bfe4fc9b
The output of image2:
https://github.com/user-attachments/assets/ac7f3573-e0cb-4d05-8c40-4358946b1d0b

Why do objects occupy different proportions for the same 512 resolution?

[FishWoWater]

You may have a look at the extracted meshes, I guess there are tiny structures scattered elsewhere in the space, which affects the object normalization. Also you can try multiple times under 512x resolution

[c1ircle]

You may have a look at the extracted meshes, I guess there are tiny structures scattered elsewhere in the space, which affects the object normalization. Also you can try multiple times under 512x resolution

I tried it a few times with 512 resolution and noticed a few exceptions: the generated results are usually not good if the input image is 512 resolution. There seems to be a black sheet underneath the mesh of the object, like in the video below, why is this the case?
https://github.com/user-attachments/assets/f412bfae-c6fa-4b88-a290-4854b4f0b97f