Very strict NSFW censor
I just implemented the sample code with the prompt "a photo of an astronaut riding a horse on mars".
Withimage = pipe(prompt, passes=20, prompt_strength=20, width=160, height=160).images[0]?
the model returns a black censored image...
Is the model too keen to cencor everything or are there something in the arguments that forces this?
I am trying to do a matrix of different parameters for reference; if this censorship threshold is this low, I think I have to rethink my tactics?
Well, I ran in to this old thread:
https://github.com/CompVis/stable-diffusion/issues/239
So it seems, it might be my graphics card, that causes the image to be totally green, which is in turn deducted to be potentially NSFW.
For me it worked to pass argument safety_checker=None to the StableDiffusionPipeline.from_pretrained(), but this of course brings a warning.
I am running RTX 4080 with 16GB. Very much appreciated, if someone would have more information on the matter, I am sensing dark clouds in my script future with this underlying behaviour.