with XL ，When the image is enlarged for viewing，The drawn image will appear as red dots in most cases

[zhaoyun0071]

Describe the bug

with XL ，When the image is enlarged for viewing，The drawn image will appear as red dots in most cases

Reproduction

from diffusers import StableDiffusionXLPipeline

base_model_path = r"sd_xl_base_1.0.safetensors"
import torch

pipe = StableDiffusionXLPipeline.from_single_file(base_model_path, torch_dtype=torch.float16,
                                                  use_safetensors=True, load_safety_checker=False).to('cuda')

image = pipe(prompt="a girl, realistic Documentary photography", guidance_scale=7, height=1280, width=768,
     num_inference_steps=30,negative_prompt="").images[0]

image.save('test.png')

Logs

No response

System Info

diffusers 0.27.0

Who can help?

@yiyixuxu @sayakpaul @DN6

[sayakpaul]

What happens when you use a different VAE like this one?

[zhaoyun0071]

What happens when you use a different VAE like this one?

still has red dots

with code bellow：

from diffusers import StableDiffusionXLPipeline
from diffusers import DiffusionPipeline, AutoencoderKL

base_model_path = r"sd_xl_base_1.0.safetensors"
import torch

#
pipe = StableDiffusionXLPipeline.from_single_file(base_model_path, torch_dtype=torch.float16,
                                                  use_safetensors=True, load_safety_checker=False,
                                                  vae=AutoencoderKL.from_single_file(r"sdxl.vae.safetensors",
                                                                                     torch_dtype=torch.float16)).to(
    'cuda')

image = pipe(prompt="a girl, realistic Documentary photography", guidance_scale=7, height=1280, width=768,
             num_inference_steps=30, negative_prompt="").images[0]

image.save('test.png')

[sayakpaul]

Is this resolution-independent? Does the same happen when you use height=width=1024?

[zhaoyun0071]

这是否与分辨率无关？当您使用 height=width=1024 时，也会发生同样的情况吗？

still has red and green dots

[bghira]

this is the watermarker, see #7013

[tolgacangoz]

When the invisible-watermark package is installed these red dots appear as a watermark. IMO, they are too visible and incredibly annoying 😞.

[zhaoyun0071]

this is the watermarker, see #7013

thanks sovle it，with add_watermarker=False

from diffusers import StableDiffusionXLPipeline
from diffusers import DiffusionPipeline, AutoencoderKL

base_model_path = r"sd_xl_base_1.0.safetensors"
import torch

#
pipe = StableDiffusionXLPipeline.from_single_file(base_model_path, torch_dtype=torch.float16,
                                                  use_safetensors=True, load_safety_checker=False,
                                                  vae=AutoencoderKL.from_single_file(r"sdxl.vae.safetensors",
                                                                                     torch_dtype=torch.float16),add_watermarker=False).to(
    'cuda')

image = pipe(prompt="a girl, realistic Documentary photography", guidance_scale=7, height=1024, width=1024,
             num_inference_steps=30, negative_prompt="").images[0]

image.save('test.png')

[sayakpaul]

Thanks, folks for the discussion. Closing the discussion but feel free to re-open.