[Do not merge] Script to compare safety checkers

[patrickvonplaten]

Open clip is taken from here: https://github.com/mlfoundations/open_clip

Script to compare the two:

from diffusers import StableDiffusionPipeline, DDIMScheduler
from time import time
from PIL import Image
from einops import rearrange
import numpy as np
import torch
from torch import autocast
from torchvision.utils import make_grid

torch.manual_seed(42)

prompts = ["a photograph of an astronaut riding a horse"]
prompts.append("a photograph of the eiffel tower on the moon")
prompts.append("an oil painting of a futuristic forest gives")

#pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-3-diffusers", use_auth_token=True)  # make sure you're logged in with `huggingface-cli login`
pipe = StableDiffusionPipeline.from_pretrained("fusing/sd-v1-3", use_auth_token=True)  # make sure you're logged in with `huggingface-cli login`

all_images = []
num_rows = 1
num_columns = 4
for prompt in prompts:
    with autocast("cuda"):
        images = pipe(num_columns * [prompt], guidance_scale=7.5, output_type="np")["sample"]  # image here is in [PIL format](https://pillow.readthedocs.io/en/stable/)
        all_images.append(torch.from_numpy(images))

# additionally, save as grid
grid = torch.stack(all_images, 0)
grid = rearrange(grid, 'n b h w c -> (n b) h w c')
grid = rearrange(grid, 'n h w c -> n c h w')
grid = make_grid(grid, nrow=num_rows)

# to image
grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
image = Image.fromarray(grid.astype(np.uint8))

image.save(f"./images/diffusers/final_{'_'.join(prompt.split())}_{round(time())}.png")

[patrickvonplaten]

This means that if adjustment > concept_threshold -> the image will always be a bad image

[yuimo]

what's the theory behind this code? why you calculate the similarity between the embedding of the image and a full-ones tensor below:
self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
thanks very much

[patrickvonplaten]

I just means that if cosine similarity is above a certain threshold then images will be blocked

[patrickvonplaten]

Safety Module is checked and works!