The compression artifacts are very inconsistent. Like generally when you get compression artifacts on an image, they’re of the same scale across the whole image, because the same process has been applied to the whole image. But with a generated image, there will be some base line compression if the image has been shared, but under that there will be weird artifacting that only appears in certain parts of the image that’s being applied during image generation due to artifacts on images in the training data.
Like, compare the text in the blue sky post, the artifacting there is very consistent, but it’s all over the place in the milk image with certain areas having massive nonsense artifacts, while other parts no more than the text in the blue sky image.
The compression artifacts are very inconsistent. Like generally when you get compression artifacts on an image, they’re of the same scale across the whole image, because the same process has been applied to the whole image. But with a generated image, there will be some base line compression if the image has been shared, but under that there will be weird artifacting that only appears in certain parts of the image that’s being applied during image generation due to artifacts on images in the training data.
Like, compare the text in the blue sky post, the artifacting there is very consistent, but it’s all over the place in the milk image with certain areas having massive nonsense artifacts, while other parts no more than the text in the blue sky image.