Red-Green dots appear on image

#6
by MagicalJourney - opened

Some kind of artifacts can be seem on higher zoom level,
the image contains thousands of red and green blobs of several pixels each.

cat_near_dots.png

Look at area left of cat.
prompt is "an image of white cat in a chair in a tea room"

I see, this is very interesting. Could you open a Discussion with the original author of the model please?

Opened a thread there:
https://huggingface.co/dataautogpt3/OpenDalleV1.1/discussions/28

I'm not sure why these artifacts appear, but they're present in lots of images(probably all but
monotonic color areas make them easier to spot). Something like anti-AI countermeasure
or watermark?

I see, this is very interesting.

@mrfakename , this is the standard "invisible watermark" in sdxl. You can (must) simply disable it in your code that uses this space. I don't understand why this isn't disabled by default...
See your requirements.txt file on the fourth line: https://huggingface.co/spaces/mrfakename/OpenDalleV1.1-GPU-Demo/blob/b8945f30a000a4914b938f4fbbc7c79b7596b8f8/requirements.txt#L4
For more details, see the put_watermark() function:
https://github.com/Stability-AI/stablediffusion/blob/cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf/scripts/txt2img.py#L206
Usually, I just edit this function by adding return img right after the function declaration

Sign up or log in to comment