from diffusers import StableDiffusionXLInpaintPipeline import gradio as gr import numpy as np import time import math import random import imageio from PIL import Image from PIL import ImageFilter import torch max_64_bit_int = 2**63 - 1 device = "cuda" if torch.cuda.is_available() else "cpu" pipe = StableDiffusionXLInpaintPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0") pipe = pipe.to(device) def noise_color(color, noise): return color + random.randint(- noise, noise) def predict( source_img, enlarge_top, enlarge_right, enlarge_bottom, enlarge_left, prompt, negative_prompt, smooth_border, denoising_steps, num_inference_steps, guidance_scale, randomize_seed, seed, debug_mode, progress = gr.Progress()): start = time.time() progress(0, desc = "Preparing data...") if source_img is None: raise gr.Error("Please provide an image.") if enlarge_top is None or enlarge_top == "": enlarge_top = 0 if enlarge_right is None or enlarge_right == "": enlarge_right = 0 if enlarge_bottom is None or enlarge_bottom == "": enlarge_bottom = 0 if enlarge_left is None or enlarge_left == "": enlarge_left = 0 if enlarge_top < 0 or enlarge_right < 0 or enlarge_bottom < 0 or enlarge_left < 0: raise gr.Error("Please only provide positive margins.") if enlarge_top == 0 and enlarge_right == 0 and enlarge_bottom == 0 and enlarge_left == 0: raise gr.Error("At least one border must be enlarged.") if prompt is None or prompt == "": raise gr.Error("Please provide a prompt input.") if negative_prompt is None: negative_prompt = "" if smooth_border is None: smooth_border = 20 if randomize_seed: seed = random.randint(0, max_64_bit_int) random.seed(seed) #pipe = pipe.manual_seed(seed) try: imageio.imwrite("data.png", source_img) except: raise gr.Error("Can't read input image. You can try to first save your image in another format (.webp, .png, .jpeg, .bmp...).") # Input image try: input_image = Image.open("data.png").convert("RGB") except: raise gr.Error("Can't open input image. You can try to first save your image in another format (.webp, .png, .jpeg, .bmp...).") original_height, original_width, original_channel = np.array(input_image).shape output_width = enlarge_left + original_width + enlarge_right output_height = enlarge_top + original_height + enlarge_bottom # Enlarged image enlarged_image = Image.new(mode = input_image.mode, size = (original_height, original_width), color = "black") enlarged_image.paste(input_image, (0, 0)) enlarged_image = enlarged_image.resize((output_width, output_height)) enlarged_image = enlarged_image.filter(ImageFilter.BoxBlur(20)) enlarged_image.paste(input_image, (enlarge_left, enlarge_top)) horizontally_mirrored_input_image = input_image.transpose(Image.FLIP_LEFT_RIGHT).resize((original_width * 2, original_height)) enlarged_image.paste(horizontally_mirrored_input_image, (enlarge_left - (original_width * 2), enlarge_top)) enlarged_image.paste(horizontally_mirrored_input_image, (enlarge_left + original_width, enlarge_top)) vertically_mirrored_input_image = input_image.transpose(Image.FLIP_TOP_BOTTOM).resize((original_width, original_height * 2)) enlarged_image.paste(vertically_mirrored_input_image, (enlarge_left, enlarge_top - (original_height * 2))) enlarged_image.paste(vertically_mirrored_input_image, (enlarge_left, enlarge_top + original_height)) returned_input_image = input_image.transpose(Image.ROTATE_180).resize((original_width * 2, original_height * 2)) enlarged_image.paste(returned_input_image, (enlarge_left - (original_width * 2), enlarge_top - (original_height * 2))) enlarged_image.paste(returned_input_image, (enlarge_left - (original_width * 2), enlarge_top + original_height)) enlarged_image.paste(returned_input_image, (enlarge_left + original_width, enlarge_top - (original_height * 2))) enlarged_image.paste(returned_input_image, (enlarge_left + original_width, enlarge_top + original_height)) enlarged_image = enlarged_image.filter(ImageFilter.BoxBlur(20)) # Noise image noise_image = Image.new(mode = input_image.mode, size = (output_width, output_height), color = "black") enlarged_pixels = enlarged_image.load() for i in range(output_width): for j in range(output_height): enlarged_pixel = enlarged_pixels[i, j] noise = min(max(enlarge_left - i, i - (enlarge_left + original_width), enlarge_top - j, j - (enlarge_top + original_height), 0), 255) noise_image.putpixel((i, j), (noise_color(enlarged_pixel[0], noise), noise_color(enlarged_pixel[1], noise), noise_color(enlarged_pixel[2], noise), 255)) enlarged_image.paste(noise_image, (0, 0)) enlarged_image.paste(input_image, (enlarge_left, enlarge_top)) # Mask mask_image = Image.new(mode = input_image.mode, size = (output_width, output_height), color = (255, 255, 255, 0)) black_mask = Image.new(mode = input_image.mode, size = (original_width - smooth_border, original_height - smooth_border), color = (127, 127, 127, 0)) mask_image.paste(black_mask, (enlarge_left + (smooth_border // 2), enlarge_top + (smooth_border // 2))) mask_image = mask_image.filter(ImageFilter.BoxBlur((smooth_border // 2))) limitation = ""; # Limited to 1 million pixels if 1024 * 1024 < output_width * output_height: factor = ((1024 * 1024) / (output_width * output_height))**0.5 output_width = math.floor(output_width * factor) output_height = math.floor(output_height * factor) limitation = " Due to technical limitation, the image have been downscaled."; # Width and height must be multiple of 8 output_width = output_width - (output_width % 8) output_height = output_height - (output_height % 8) progress(None, desc = "Processing...") output_image = pipe(seeds=[seed], width = output_width, height = output_height, prompt = prompt, negative_prompt = negative_prompt, image = enlarged_image, mask_image = mask_image, num_inference_steps = num_inference_steps, guidance_scale = guidance_scale, denoising_steps = denoising_steps, show_progress_bar = True).images[0] if debug_mode == False: input_image = None enlarged_image = None mask_image = None end = time.time() return [ output_image, "Start again to get a different result. The new image is " + str(output_width) + " pixels large and " + str(output_height) + " pixels high, so an image of " + str(output_width * output_height) + " pixels. The image have been generated in " + str(end - start) + " seconds." + limitation, input_image, enlarged_image, mask_image ] def toggle_debug(is_debug_mode): if is_debug_mode: return [gr.update(visible = True)] * 3 else: return [gr.update(visible = False)] * 3 with gr.Blocks() as interface: gr.Markdown( """
Uncrop
Enlarges the point of view of your image, up to 1 million pixels, freely, without account, without watermark, which can be downloaded