Spaces:
Running
on
A10G
Running
on
A10G
File size: 3,375 Bytes
9c0622a 383b0d8 9c0622a 383b0d8 9c0622a 383b0d8 b08b4dd 383b0d8 b08b4dd 383b0d8 0b6171d 383b0d8 69ec261 383b0d8 69ec261 383b0d8 b32b037 383b0d8 0b6171d 383b0d8 b32b037 67f8600 b32b037 383b0d8 b32b037 383b0d8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 |
import gradio as gr
import torch
from PIL import Image
from lambda_diffusers import StableDiffusionImageEmbedPipeline
def main(
input_im,
scale=3.0,
n_samples=4,
steps=25,
seed=0,
):
generator = torch.Generator(device=device).manual_seed(int(seed))
images_list = pipe(
n_samples*[input_im],
guidance_scale=scale,
num_inference_steps=steps,
generator=generator,
)
images = []
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
description = \
"""
__This demo is currently running on CPU only, so takes a looooooong time, a GPU will be added in the near future!
In the meantime test out the version [here](https://47725.gradio.app/).__
Generate variations on an input image using a fine-tuned version of Stable Diffision.
Trained by [Justin Pinkney](https://www.justinpinkney.com) ([@Buntworthy](https://twitter.com/Buntworthy)) at [Lambda](https://lambdalabs.com/)
This version has been ported to 🤗 Diffusers library, see more details on how to use this version in the [Lambda Diffusers repo](https://github.com/LambdaLabsML/lambda-diffusers).
__For the original training code see [this repo](https://github.com/justinpinkney/stable-diffusion).
"""
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.
"""
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, 25, value=3, step=1, label="Guidance scale"),
gr.Slider(1, 4, value=1, step=1, label="Number images"),
gr.Slider(5, 50, value=25, step=5, label="Steps"),
gr.Number(0, labal="Seed", precision=0)
]
output = gr.Gallery(label="Generated variations")
output.style(grid=2)
examples = [
["examples/vermeer.jpg", 3, 1, 25, 0],
["examples/matisse.jpg", 3, 1, 25, 0],
]
demo = gr.Interface(
fn=main,
title="Stable Diffusion Image Variations",
description=description,
article=article,
inputs=inputs,
outputs=output,
examples=examples,
)
demo.launch()
|