Disabling safety-model or fixing false positives?

[bartekleon]

I really have wanted to try this project so i tried using it with diffusers (default configurion for this one throws lack of memory and that one actually runs). I am getting "Potential NSFW content was detected in one or more images. A black image will be returned instead. Try again with a different prompt and/or seed." errpr all the time. With a default prompt. Or a "star" if "riding horse is NSFW". Is it possible to debug the runs it so the "handmade" safety feature is fixed or checked?
Or is there possibility to run this without this safety model or with it switched off?
Running default

import torch
from torch import autocast
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "./stable-diffusion-v1-4",
    revision="fp16", 
    torch_dtype=torch.float16
)
pipe = pipe.to("cuda")

prompt = "star"
with autocast("cuda"):
    image = pipe(prompt).images[0]  
image.save("star.png")

(tried seed 12345 if someone wants to try reproduce. Maybe my PC is not safe for work or something)
Thanks in advance

[tails101]

i agree a nsfw toggle would be nice, sometime just asking for a swimsuit trigger it.

[rogeriochaves]

disabling it is easy, you can do this:

pipe.safety_checker = lambda images, clip_input: (images, False)

[peterbayerle]

I also found the checker was inaccurately flagging some of my prompts as NSFW, so I disabled it in scripts/txt2img.py, by replacing check_safety with:

def check_safety(x_image):
    # safety_checker_input = safety_feature_extractor(numpy_to_pil(x_image), return_tensors="pt")
    # x_checked_image, has_nsfw_concept = safety_checker(images=x_image, clip_input=safety_checker_input.pixel_values)
    # assert x_checked_image.shape[0] == len(has_nsfw_concept)
    # for i in range(len(has_nsfw_concept)):
    #     if has_nsfw_concept[i]:
    #         x_checked_image[i] = load_replacement(x_checked_image[i])
    # return x_checked_image, has_nsfw_concept
    return x_image, False

[bartekleon]

It seems my issue was the problem with GTX 1660 and numpy with fp16. So basically it was making these all green images, which were later marked by NSFW filter as NSFW and turned them black. It still is strange behaviour, but using the low memory fork with the higher quality fixed it

[TheConceptBoy]

disabling it is easy, you can do this:

pipe.safety_checker = lambda images, clip_input: (images, False)

All this did was disable the message, the image still came out black.

[TheConceptBoy]

It seems my issue was the problem with GTX 1660 and numpy with fp16. So basically it was making these all green images, which were later marked by NSFW filter as NSFW and turned them black. It still is strange behaviour, but using the low memory fork with the higher quality fixed it

Could you share the link to the low memory fork? The issue persists with a GTX1650 as well.

[bartekleon]

It seems my issue was the problem with GTX 1660 and numpy with fp16. So basically it was making these all green images, which were later marked by NSFW filter as NSFW and turned them black. It still is strange behaviour, but using the low memory fork with the higher quality fixed it

Could you share the link to the low memory fork? The issue persists with a GTX1650 as well.

https://github.com/basujindal/stable-diffusion
It was this one. I had to add "--precision full" tho in order for it to work correctly. Although it does increase VRAM it is the only thing to make it work
"If you have a Nvidia GTX series GPU, the output images maybe entirely green in color. This is because GTX series do not support half precision calculation, which is the default mode of calculation in this repository. To overcome the issue, use the --precision full argument. The downside is that it will lead to higher GPU VRAM usage."
It should take ~3GB VRAM for 512x512

[Nukami]

disabling it is easy, you can do this:

pipe.safety_checker = lambda images, clip_input: (images, False)

All this did was disable the message, the image still came out black.

It actually worked on me. This line overwrites the safety_checker of StableDiffusionPipeline by returning the origin image, graceful and efficacious.

Just overwrite the safety_checker after where a StableDiffusionPipeline is initialized

[andresberejnoi]

i agree a nsfw toggle would be nice, sometime just asking for a swimsuit trigger it.

I added a simple toggle for the txt2img.py script. #442

[polym]

disabling it is easy, you can do this:

pipe.safety_checker = lambda images, clip_input: (images, False)

Please note that this should be placed after calling pipe.enable_model_cpu_offload(). Otherwise, you may encounter an exception with the message "AttributeError: 'function' object has no attribute 'forward'".

[JemiloII]

Just was looking through the code and found that we can set it to this without requiring a lambda. Though in the future, if they change this, then the lambda will probably be necessary.

StableDiffusionPipeline.from_pretrained(
    "./stable-diffusion-v1-4",
    revision="fp16", 
    torch_dtype=torch.float16
    safety_checker = None,
    requires_safety_checker = False
)

You can also change it later if necessary by doing this.

pipe.safety_checker = None
pipe.requires_safety_checker = False

[1318980306]

It seems my issue was the problem with GTX 1660 and numpy with fp16. So basically it was making these all green images, which were later marked by NSFW filter as NSFW and turned them black. It still is strange behaviour, but using the low memory fork with the higher quality fixed it

Could you share the link to the low memory fork? The issue persists with a GTX1650 as well.

https://github.com/basujindal/stable-diffusion It was this one. I had to add "--precision full" tho in order for it to work correctly. Although it does increase VRAM it is the only thing to make it work "If you have a Nvidia GTX series GPU, the output images maybe entirely green in color. This is because GTX series do not support half precision calculation, which is the default mode of calculation in this repository. To overcome the issue, use the --precision full argument. The downside is that it will lead to higher GPU VRAM usage." It should take ~3GB VRAM for 512x512

Sorry, can you mention how excatly what to do, I couldn't add "--precision full" this agrument. it will poop up syntaxerror: invaild syntax. Please help me.

[JemiloII]

@1318980306 Your graphics card is too old. I had to upgrade mine in order to use full.