Watermark bias about stablediffusion

[gonferzika]

I get some images with watermarks here too, and i just begin to play with the new model, the previous V1 almost never generate watermarks for me.

[yunnieXx]

Pretty sure you can use a negative tag to remove most watermarks (watermark:-1), but yes, they probably trained 2.0 on more stock images with watermarks that before.

[maciej-zuk]

@yunnieXx I have tried negative prompts and it doesn't fix the problem. CompVis' SD model card (https://huggingface.co/CompVis/stable-diffusion) mentions they filtered out images with watermark probability > 50%, I wonder if authors of SD2 applied same filter to training data.