Fix AttributeError of VisualClozeProcessor

[Justin900429]

Summary

Fixes an AttributeError in VisualClozePipeline where the code attempted to access a non-existent height function on VisualClozeProcessor. The pipeline now calls the correct resizing utility (_resize_and_crop) during preprocessing.

diffusers/src/diffusers/pipelines/visualcloze/visualcloze_utils.py

Lines 105 to 113 in f442955

	if len(target_position) > 1 and sum(target_position) > 1:
	new_w = resize_size[n_samples - 1][0] or 384
	for i in range(len(processed_images)):
	for j in range(len(processed_images[i])):
	if processed_images[i][j] is not None:
	new_h = int(processed_images[i][j].height * (new_w / processed_images[i][j].width))
	new_w = int(new_w / 16) * 16
	new_h = int(new_h / 16) * 16
	processed_images[i][j] = self.height(processed_images[i][j], new_h, new_w)

This error occurs only when generating more than one image.

Reproduction

from diffusers import VisualClozePipeline
from PIL import Image
import torch

image_paths = [
    [
        Image.new("RGB", (384, 384), (0, 0, 0)),
        Image.new("RGB", (384, 384), (0, 0, 0)),
        Image.new("RGB", (384, 384), (0, 0, 0)),
    ],
    [
        Image.new("RGB", (384, 384), (0, 0, 0)),
        None,
        None,
    ],
]

task_prompt = "test"
content_prompt = None

pipe = VisualClozePipeline.from_pretrained(
    "VisualCloze/VisualClozePipeline-384", resolution=384, torch_dtype=torch.bfloat16
).to("cuda")

image_result = pipe(
    task_prompt=task_prompt,
    content_prompt=content_prompt,
    image=image_paths,
    upsampling_width=512,
    upsampling_height=512,
    upsampling_strength=0.0,
    guidance_scale=30,
    num_inference_steps=30,
    max_sequence_length=512,
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]

• Error:

AttributeError: 'VisualClozeProcessor' object has no attribute 'height'

@yiyixuxu @asomoza

[a-r-r-o-w]

Looks correct to me! Just curious, why calling self._resize_and_crop and not self.resize?

[Justin900429]

Thanks for the reply!

Not sure which one is the author’s intended approach, but since they use _resize_and_crop above for the same function, I just followed their implementation.

Reference:

diffusers/src/diffusers/pipelines/visualcloze/visualcloze_utils.py

Line 94 in 03be15e

target = self._resize_and_crop(input_images[i][j], resize_size[i][0], resize_size[i][1])

Edit:

In the authors’ original repo, they apply resize first and then perform a center crop. Therefore, using _resize_and_crop better aligns with their original implementation. (Check here)