Safety Filter

[jochemstoel]

When I use this template to deploy on your platform and do inference on Stable Diffusion, how do I get rid of the safety filter check? (check_safety in txt2img.py) This is a requirement and one of the reasons to deploy myself.

[gadicc]

In your app.py, define:

class DummySafetyChecker:
    @staticmethod
    def __call__(images, clip_input):
        return images, False

and then use it when defining your model:

def init():
  # ...
  model = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    scheduler=lms,
    safety_checker=DummySafetyChecker()  # <-- add this
    use_auth_token=HF_AUTH_TOKEN
  ).to("cuda")