this repo contains the default image preprocessing code for wan 2.2 5B I2V pipeline
this will load an image from URL, resize and center crop it
image_processor = ModularPipeline.from_pretrained("YiYiXu/WanImageProcessor", trust_remote_code=True)
image = image_processor(
image="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG",
max_area=1280*704, output="processed_image")
it does this:
# copied from https://github.com/Wan-Video/Wan2.2/blob/388807310646ed5f318a99f8e8d9ad28c5b65373/wan/utils/utils.py#L136
def best_output_size(w, h, dw, dh, expected_area):
# float output size
ratio = w / h
ow = (expected_area * ratio)**0.5
oh = expected_area / ow
# process width first
ow1 = int(ow // dw * dw)
oh1 = int(expected_area / ow1 // dh * dh)
assert ow1 % dw == 0 and oh1 % dh == 0 and ow1 * oh1 <= expected_area
ratio1 = ow1 / oh1
# process height first
oh2 = int(oh // dh * dh)
ow2 = int(expected_area / oh2 // dw * dw)
assert oh2 % dh == 0 and ow2 % dw == 0 and ow2 * oh2 <= expected_area
ratio2 = ow2 / oh2
# compare ratios
if max(ratio / ratio1, ratio1 / ratio) < max(ratio / ratio2,
ratio2 / ratio):
return ow1, oh1
else:
return ow2, oh2
from diffusers.utils import load_image
image = load_image(block_state.image).convert("RGB")
max_area = 1280*704
ih, iw = image.height, image.width
dh, dw = pipe.transformer.config.patch_size[1] * pipe.vae_scale_factor_spatial, pipe.transformer.config.patch_size[2] * pipe.vae_scale_factor_spatial
ow, oh = best_output_size(iw, ih, dw, dh, max_area)
scale = max(ow / iw, oh / ih)
resized_image = image.resize((round(iw * scale), round(ih * scale)), Image.LANCZOS)
# center-crop
x1 = (resized_image.width - ow) // 2
y1 = (resized_image.height - oh) // 2
image = resized_image.crop((x1, y1, x1 + ow, y1 + oh))