app.py

import gradio as gr
import torch
from PIL import Image

from lambda_diffusers import StableDiffusionImageEmbedPipeline

def ask(input_im, scale, steps, seed, images):
    images = images
    generator = torch.Generator(device=device).manual_seed(int(seed))
    
    images_list = pipe(
        2*[input_im],
        guidance_scale=scale,
        num_inference_steps=steps,
        generator=generator,
        )
    
    for i, image in enumerate(images_list["sample"]):
        if(images_list["nsfw_content_detected"][i]):
            safe_image = Image.open(r"unsafe.png")
            images.append(safe_image)
        else:
            images.append(image)
    return images

def main(input_im, n_pairs, scale, steps, seed):
    print('Start the magic !')
    images = []
    for i in range(n_pairs):
        print('Asking for a new pair of image [' + str(i + 1) + '/' + str(n_pairs) + ']')
        seed = seed+i
        images = ask(input_im, scale, steps, seed, images)
    print('Thanks to Sylvain, it worked like a charm!')
    return images

device = "cuda" if torch.cuda.is_available() else "cpu"
pipe = StableDiffusionImageEmbedPipeline.from_pretrained(
    "lambdalabs/sd-image-variations-diffusers",
    revision="273115e88df42350019ef4d628265b8c29ef4af5",
    )
pipe = pipe.to(device)

inputs = [
    gr.Image(),
    gr.Slider(0, 3, value=2, step=1, label="Pairs of images to ask"),
    gr.Slider(0, 25, value=3, step=1, label="Guidance scale"),
    gr.Slider(5, 50, value=25, step=5, label="Steps"),
    gr.Slider(label = "Seed", minimum = 0, maximum = 2147483647, step = 1, randomize = True)
]
output = gr.Gallery(label="Generated variations")
output.style(grid=2, height="")

description = \
"""
<p style='text-align: center;'>This demo is running on CPU. Working version fixed by Sylvain <a href='https://twitter.com/fffiloni' target='_blank'>@fffiloni</a>. You'll get n pairs of images variations. <br /> 
Asking for pairs of images instead of more than 2 images in a row helps us to avoid heavy CPU load and connection error out ;)<br />
Waiting time (for 2 pairs): ~5/10 minutes • NSFW filters enabled • <img id='visitor-badge' alt='visitor badge' src='https://visitor-badge.glitch.me/badge?page_id=gradio-blocks.sd-img-variations' style='display: inline-block' /><br />
Generate variations on an input image using a fine-tuned version of Stable Diffusion.<br />
Trained by <a href='https://www.justinpinkney.com' target='_blank'>Justin Pinkney</a> (<a href='https://twitter.com/Buntworthy' target='_blank'>@Buntworthy</a>) at <a href='https://lambdalabs.com/' target='_blank'>Lambda</a><br />
This version has been ported to 🤗 Diffusers library, see more details on how to use this version in the <a href='https://github.com/LambdaLabsML/lambda-diffusers' target='_blank'>Lambda Diffusers repo</a>.<br />
For the original training code see <a href='https://github.com/justinpinkney/stable-diffusion' target='_blank'>this repo</a>.
<img src='https://raw.githubusercontent.com/justinpinkney/stable-diffusion/main/assets/im-vars-thin.jpg' style='display: inline-block;' />
</p>
"""

article = \
"""
—
## How does this work?
The normal Stable Diffusion model is trained to be conditioned on text input. This version has had the original text encoder (from CLIP) removed, and replaced with
the CLIP _image_ encoder instead. So instead of generating images based a text input, images are generated to match CLIP's embedding of the image.
This creates images which have the same rough style and content, but different details, in particular the composition is generally quite different.
This is a totally different approach to the img2img script of the original Stable Diffusion and gives very different results.
The model was fine tuned on the [LAION aethetics v2 6+ dataset](https://laion.ai/blog/laion-aesthetics/) to accept the new conditioning.
Training was done on 4xA6000 GPUs on [Lambda GPU Cloud](https://lambdalabs.com/service/gpu-cloud).
More details on the method and training will come in a future blog post.
"""

demo = gr.Interface(
    fn=main,
    title="Stable Diffusion Image Variations",
    inputs=inputs,
    outputs=output,
    description=description,
    article=article
    )
demo.launch()