[Training] Better image interpolation in training scripts

[asomoza]

What does this PR do?

As discussed here, this PR will add LANCZOS as a default interpolation mode for the image resizing in the training scripts and if the users prefers can choose BILINEAR.

I'll add this to the most popular recent ones and leave the rest to the community if they want to add them to other training scripts.

I'll do some training runs first to test if I can see the difference, but still, I already know that LANCZOS is better and that the models can pick subtle details that the human eye can't.

Fixes #6397

Who can review?

@bghira @linoytsaban @sayakpaul

Additional

Here's a little script if you want to test and try to see the difference:

import matplotlib.pyplot as plt
from PIL import Image
from torchvision import transforms

from diffusers.utils import load_image


image = load_image("https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/training/circle.png")

# Define the resize transforms for 256x256
resize_bilinear_256 = transforms.Resize((256, 256), interpolation=transforms.InterpolationMode.BILINEAR)
resize_lanczos_256 = transforms.Resize((256, 256), interpolation=transforms.InterpolationMode.LANCZOS)

# Apply the 256x256 transforms
image_bilinear_256 = resize_bilinear_256(image)
image_lanczos_256 = resize_lanczos_256(image)
image_pil_lanczos_256 = image.resize((256, 256), Image.Resampling.LANCZOS)

# Define the resize transforms for 32x32
resize_bilinear_32 = transforms.Resize((32, 32), interpolation=transforms.InterpolationMode.BILINEAR)
resize_lanczos_32 = transforms.Resize((32, 32), interpolation=transforms.InterpolationMode.LANCZOS)

# Apply the 32x32 transforms
image_bilinear_32 = resize_bilinear_32(image)
image_lanczos_32 = resize_lanczos_32(image)
image_pil_lanczos_32 = image.resize((32, 32), Image.Resampling.LANCZOS)

# Display the results
plt.figure(figsize=(16, 12))

# First row - 256x256 resizing
plt.subplot(2, 4, 1)
plt.title("Original Image")
plt.imshow(image)
plt.axis("off")

plt.subplot(2, 4, 2)
plt.title("Bilinear (256x256)")
plt.imshow(image_bilinear_256)
plt.axis("off")

plt.subplot(2, 4, 3)
plt.title("torchvision Lanczos (256x256)")
plt.imshow(image_lanczos_256)
plt.axis("off")

plt.subplot(2, 4, 4)
plt.title("PIL Lanczos (256x256)")
plt.imshow(image_pil_lanczos_256)
plt.axis("off")

# Second row - 32x32 resizing
plt.subplot(2, 4, 5)
plt.title("Original Image")
plt.imshow(image)
plt.axis("off")

plt.subplot(2, 4, 6)
plt.title("Bilinear (32x32)")
plt.imshow(image_bilinear_32)
plt.axis("off")

plt.subplot(2, 4, 7)
plt.title("torchvision Lanczos (32x32)")
plt.imshow(image_lanczos_32)
plt.axis("off")

plt.subplot(2, 4, 8)
plt.title("PIL Lanczos (32x32)")
plt.imshow(image_pil_lanczos_32)
plt.axis("off")

plt.tight_layout()
plt.show()

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[bghira]

thank you

[sayakpaul]

LGTM after the comments are resolved.

[sayakpaul]

@asomoza hope this is okay.

[sayakpaul]

Failing test is unrelated.