| 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, scale, steps, seed): | |
| images = [] | |
| images = ask(input_im, scale, n_samples, steps, seed, images) | |
| images = ask(input_im, scale, n_samples, steps, seed, images) | |
| 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, 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 @fffiloni. You'll get 4 images variations. 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. | |
| 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).__ | |
|  | |
| </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() | |