Any reason invisible watermarking isn't included and on by default?

Hi all,

Love the work y’all are doing!

The original CompVis Stable Diffusion repo includes invisible watermarking so that images generated don’t end-up polluting newly scraped training data with automatically generated results, as is already happening for text data..

I would be happy to start an issue on the repo, but I wanted to check here first to see if this somehow already happening under the hood in a way I couldn’t find, or if this had already been discussed?


1 Like

Welcome @ari-holtzman! That’s a good question! I’m not quite sure whether the team has discussed including the invisible watermarking yet, but I’ll pass this by them. In the meantime, if you’re interested in opening up an issue on the repo (GitHub - huggingface/diffusers: 🤗 Diffusers: State-of-the-art diffusion models for image and audio generation in PyTorch) that would be great!