import gradio as gr from gradio_imageslider import ImageSlider import torch from diffusers import DiffusionPipeline, AutoencoderKL, ControlNetModel from compel import Compel, ReturnedEmbeddingsType from PIL import Image from torchvision import transforms import tempfile import os import time import uuid import cv2 import numpy as np device = "cuda" if torch.cuda.is_available() else "cpu" dtype = torch.float16 LOW_MEMORY = os.getenv("LOW_MEMORY", "0") == "1" print(f"device: {device}") print(f"dtype: {dtype}") print(f"low memory: {LOW_MEMORY}") vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype) controlnet = ControlNetModel.from_pretrained( "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16 ) pipe = DiffusionPipeline.from_pretrained( "stabilityai/stable-diffusion-xl-base-1.0", custom_pipeline="pipeline_demofusion_sdxl_controlnet.py", controlnet=controlnet, custom_revision="main", torch_dtype=dtype, variant="fp16", use_safetensors=True, vae=vae, ) compel = Compel( tokenizer=[pipe.tokenizer, pipe.tokenizer_2], text_encoder=[pipe.text_encoder, pipe.text_encoder_2], returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, requires_pooled=[False, True], ) pipe = pipe.to(device) def load_and_process_image(pil_image): transform = transforms.Compose( [ transforms.Resize((1024, 1024)), transforms.ToTensor(), transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), ] ) image = transform(pil_image) image = image.unsqueeze(0).half() return image def pad_image(image): w, h = image.size if w == h: return image elif w > h: new_image = Image.new(image.mode, (w, w), (0, 0, 0)) pad_w = 0 pad_h = (w - h) // 2 new_image.paste(image, (0, pad_h)) return new_image else: new_image = Image.new(image.mode, (h, h), (0, 0, 0)) pad_w = (h - w) // 2 pad_h = 0 new_image.paste(image, (pad_w, 0)) return new_image def predict( input_image, prompt, negative_prompt, seed, controlnet_conditioning_scale, guidance_scale=8.5, cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8, scale=2, progress=gr.Progress(track_tqdm=True), ): if input_image is None: raise gr.Error("Please upload an image.") padded_image = pad_image(input_image).resize((1024, 1024)).convert("RGB") image_lr = load_and_process_image(padded_image).to(device) conditioning, pooled = compel([prompt, negative_prompt]) generator = torch.manual_seed(seed) last_time = time.time() canny_image = np.array(padded_image) canny_image = cv2.Canny(canny_image, 100, 200) canny_image = canny_image[:, :, None] canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2) canny_image = Image.fromarray(canny_image) images = pipe( prompt_embeds=conditioning[0:1], pooled_prompt_embeds=pooled[0:1], negative_prompt_embeds=conditioning[1:2], negative_pooled_prompt_embeds=pooled[1:2], image_lr=image_lr, width=1024 * scale, height=1024 * scale, view_batch_size=16, controlnet_conditioning_scale=controlnet_conditioning_scale, condition_image=canny_image, stride=64, generator=generator, num_inference_steps=40, guidance_scale=guidance_scale, cosine_scale_1=cosine_scale_1, cosine_scale_2=cosine_scale_2, cosine_scale_3=cosine_scale_3, sigma=sigma, multi_decoder=1024 * scale > 2048, show_image=False, lowvram=LOW_MEMORY, ) print(f"Time taken: {time.time() - last_time}") images_path = tempfile.mkdtemp() paths = [] uuid_name = uuid.uuid4() for i, img in enumerate(images): img.save(images_path + f"/img_{uuid_name}_{img.size[0]}.jpg") paths.append(images_path + f"/img_{uuid_name}_{img.size[0]}.jpg") return (images[0], images[-1]), paths css = """ #intro{ max-width: 32rem; text-align: center; margin: 0 auto; } """ with gr.Blocks(css=css) as demo: gr.Markdown( """ # Enhance This ### DemoFusion SDXL [DemoFusion](https://ruoyidu.github.io/demofusion/demofusion.html) enables higher-resolution image generation. You can upload an initial image and prompt to generate an enhanced version. [Duplicate Space](https://huggingface.co/spaces/radames/Enhance-This-DemoFusion-SDXL?duplicate=true) to avoid the queue. GPU Time Comparison: T4: ~276s - A10G: ~113.6s A100: ~43.5s RTX 4090: ~48.1s Notes The author advises against the term "super resolution" because it's more like image-to-image generation than enhancement, but it's still a lot of fun! """, elem_id="intro", ) with gr.Row(): with gr.Column(scale=1): image_input = gr.Image(type="pil", label="Input Image") prompt = gr.Textbox( label="Prompt", info="The prompt is very important to get the desired results. Please try to describe the image as best as you can. Accepts Compel Syntax", ) negative_prompt = gr.Textbox( label="Negative Prompt", value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic", ) seed = gr.Slider( minimum=0, maximum=2**64 - 1, value=1415926535897932, step=1, label="Seed", randomize=True, ) with gr.Accordion(label="DemoFusion Params", open=False): guidance_scale = gr.Slider( minimum=0, maximum=50, value=8.5, step=0.001, label="Guidance Scale", ) scale = gr.Slider( minimum=1, maximum=5, value=2, step=1, label="Magnification Scale", interactive=False, ) cosine_scale_1 = gr.Slider( minimum=0, maximum=5, value=3, step=0.01, label="Cosine Scale 1", ) cosine_scale_2 = gr.Slider( minimum=0, maximum=5, value=1, step=0.01, label="Cosine Scale 2", ) cosine_scale_3 = gr.Slider( minimum=0, maximum=5, value=1, step=0.01, label="Cosine Scale 3", ) sigma = gr.Slider( minimum=0, maximum=1, value=0.8, step=0.01, label="Sigma", ) with gr.Accordion(label="ControlNet Params", open=False): controlnet_conditioning_scale = gr.Slider( minimum=0, maximum=1, step=0.001, value=0.5, label="ControlNet Conditioning Scale", ) controlnet_start = gr.Slider( minimum=0, maximum=1, step=0.001, value=0.0, label="ControlNet Start", ) controlnet_end = gr.Slider( minimum=0.0, maximum=1.0, step=0.001, value=1.0, label="ControlNet End", ) btn = gr.Button() with gr.Column(scale=2): image_slider = ImageSlider(position=0.5) files = gr.Files() inputs = [ image_input, prompt, negative_prompt, seed, controlnet_conditioning_scale, guidance_scale, cosine_scale_1, cosine_scale_2, cosine_scale_3, sigma, # scale, ] outputs = [image_slider, files] btn.click(predict, inputs=inputs, outputs=outputs, concurrency_limit=1) gr.Examples( fn=predict, examples=[ [ "./examples/lara.jpeg", "photography of lara croft 8k high definition award winning", "blurry, ugly, duplicate, poorly drawn, deformed, mosaic", 5436236241, 0.5, 8.5, 3, 1, 1, 0.8, 2, ], [ "./examples/cybetruck.jpeg", "photo of tesla cybertruck futuristic car 8k high definition on a sand dune in mars, future", "blurry, ugly, duplicate, poorly drawn, deformed, mosaic", 383472451451, 0.5, 8.5, 3, 1, 1, 0.8, 2, ], [ "./examples/jesus.png", "a photorealistic painting of Jesus Christ, 4k high definition", "blurry, ugly, duplicate, poorly drawn, deformed, mosaic", 13317204146129588000, 0.5, 8.5, 3, 1, 1, 0.8, 2, ], [ "./examples/anna-sullivan-DioLM8ViiO8-unsplash.jpg", "A crowded stadium with enthusiastic fans watching a daytime sporting event, the stands filled with colorful attire and the sun casting a warm glow", "blurry, ugly, duplicate, poorly drawn, deformed, mosaic", 5623124123512, 0.5, 8.5, 3, 1, 1, 0.8, 2, ], [ "./examples/img_aef651cb-2919-499d-aa49-6d4e2e21a56e_1024.jpg", "a large red flower on a black background 4k high definition", "blurry, ugly, duplicate, poorly drawn, deformed, mosaic", 23123412341234, 0.5, 8.5, 3, 1, 1, 0.8, 2, ], [ "./examples/huggingface.jpg", "photo realistic huggingface human+++ emoji costume, round, yellow, skin+++ texture+++", "blurry, ugly, duplicate, poorly drawn, deformed, mosaic, emoji cartoon, drawing, pixelated", 5532144938416372000, 0.101, 25.206, 4.64, 1, 1, 0.49, 3, ], ], inputs=inputs, outputs=outputs, cache_examples=True, ) demo.queue(api_open=False) demo.launch(show_api=False)