IP adapter output on some resolutions suffers in quality?

### Describe the bug

I am running IP adapter for 768x1344 which is one of the sdxl listed resolutions. I find that the output quality is much less than say regular 768x768 generations. I've attached sample images and code below. In this experiment 1080x768 seemed to get best output, but its not one of the supported resolutions @asomo


### Reproduction

import torch
from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, ControlNetModel, StableDiffusionXLControlNetPipeline, AutoencoderKL, UniPCMultistepScheduler
from diffusers.image_processor import IPAdapterMaskProcessor
from transformers import CLIPVisionModelWithProjection
from controlnet_aux import AnylineDetector
import cv2
import numpy as np
from PIL import Image, ImageOps
from huggingface_hub import hf_hub_download

def create_controlnet_pipes(image_encoder=None)->StableDiffusionXLControlNetPipeline:
## get controlnet
controlnet = ControlNetModel.from_pretrained(
"diffusers/controlnet-canny-sdxl-1.0",
torch_dtype=torch.float16,
use_safetensors=True,
)
pipe = StableDiffusionXLPipeline.from_single_file(
"sdxl model path",
add_watermarker=False,
torch_dtype=torch.float16,
variant="fp16",
use_safetensors=True,
image_encoder=image_encoder,
)
pipe = StableDiffusionXLControlNetPipeline(
controlnet=controlnet,
**pipe.components,
add_watermarker=False,
)
pipe = pipe.to("cuda")
return pipe

def canny(image):
image = np.array(image)
low_threshold = 100
high_threshold = 200
image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
return Image.fromarray(image)

if __name__ == '__main__':
## crop different values like 0,0,1080,768 or 0,0,1280,768
ref_image = Image.open('images/fridge_fg.png').crop((0,0,1344,768))
bg_ref_image = Image.open('images/fridge_bg.png').crop((0,0,1344,768))

mask_new = Image.open('images/fridge_mask.png').convert('L').crop((0,0,1344,768))
inv_mask = Image.open('images/fridge_inv_mask.png').convert('L').crop((0,0,1344,768))
processor = IPAdapterMaskProcessor()
mask_fg = processor.preprocess([mask_new])
mask_fg = mask_fg.reshape(1, mask_fg.shape[0], mask_fg.shape[2], mask_fg.shape[3])

mask_bg = processor.preprocess([inv_mask])
mask_bg = mask_bg.reshape(1, mask_bg.shape[0], mask_bg.shape[2], mask_bg.shape[3])

canny_pil = Image.open('images/fridge_canny.png').crop((0,0,1344,768))

image_encoder = CLIPVisionModelWithProjection.from_pretrained(
"h94/IP-Adapter",
subfolder="models/image_encoder",
torch_dtype=torch.float16
)
pipe = create_controlnet_pipes(image_encoder=image_encoder)
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name=["ip-adapter-plus_sdxl_vit-h.safetensors", "ip-adapter-plus_sdxl_vit-h.safetensors"], use_safetensors=True)
scale_config_fg = {'down':1, 'mid':1, 'up':1}
scale_config_bg = {"down":0.7, 'mid':0.7, 'up':0.7}
pipe.set_ip_adapter_scale([scale_config_fg, scale_config_bg])

for idx in range(5):
outputs = pipe(
prompt='kitchen scene',
image=canny_pil,
ip_adapter_image=[ref_image, bg_ref_image],
negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality, fuzzy, blurry",
guidance_scale=5,
num_inference_steps=30,
controlnet_conditioning_scale=0.53,
cross_attention_kwargs={"ip_adapter_masks": [mask_fg, mask_bg]},
num_images_per_prompt=1
# generator=generator,
).images
for image in outputs:
image.save()
# image.save(f'output_plus/fridge_ar_ctrlnet_1280_plus_{idx}.png')
print('done')
pipe.unload_ip_adapter()

### Logs

_No response_

### System Info

v0.28.2 diffusers

### Who can help?

_No response_

Start with the provided Python reproduction using StableDiffusionXLControlNetPipeline, IPAdapterMaskProcessor, and the listed SDXL resolutions. Compare the 768x1344, 768x768, and 1080x768 runs, then trace resolution and mask handling to determine whether the quality difference is expected or a bug. Done means the behavior is explained and a reproducible expected result is established.

Vom Indexierungsmodell aus dem Issue-Text verfasst.