How to disable safety checker

[MaxTran96]

Is your feature request related to a problem? Please describe.
Yes, sometimes when i run inference, it generate an all black photos. I suspected this is due to the safety checker

Describe the solution you'd like
An option to disable safety checker

Describe alternatives you've considered
I dont know any other alternative

Additional context
N/A

[stpg06]

You pass "safety_checker=None" through the pipeline when you call the model. It's very simple.

Example from diffusers inference code might read similar to this:
'pipe = StableDiffusionPipeline.from_pretrained(model_path, safety_checker=None, torch_dtype=torch.float16).to("cuda")'

[sayakpaul]

Thanks @stpg06 for your answer.

Closing the issue, hence :)

[jtoy]

I'm using diffusers 0.16.1 (latest) and when I use :

pipe = StableDiffusionPipeline.from_ckpt(full_model_path,safety_checker=None, torch_dtype=torch.float16).to("cuda")

I am still getting black images. is there a way to confirm the safety checker is off? I just want to get rid of black images.
In older versions of diffusers, we could pass:
pipe.safety_checker = lambda images, **kwargs: (images, False)
but it seems to have broken in the later versions.

[jtoy]

I did more testing, I don't believe passing safety_checker=None works for all the pipelines (from_ckpt/from_pretrained). But passing this to a pipe works now (where it didnt work in older versions of diffusers):
pipe.safety_checker = None
So I think this ticket is a valid ticket or should be reopened as a bug.

[sayakpaul]

Could you provide a Colab Notebook that reproduces this issue?

If we can reproduce the issue there, more than happy to reopen this :)

[KennethTrinh]

I did this:

def disabled_safety_checker(images, clip_input):
    if len(images.shape)==4:
        num_images = images.shape[0]
        return images, [False]*num_images
    else:
        return images, False
pipe.safety_checker = disabled_safety_checker

Kudos to CompVis/stable-diffusion#331 (comment)