import gradio as gr import torch from diffusers import AutoencoderKL from diffusers.utils import load_image from controlnet_flux import FluxControlNetModel from transformer_flux import FluxTransformer2DModel from pipeline_flux_controlnet_inpaint import FluxControlNetInpaintingPipeline from PIL import Image, ImageDraw import numpy as np import spaces from huggingface_hub import hf_hub_download # Load models controlnet = FluxControlNetModel.from_pretrained("alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha", torch_dtype=torch.bfloat16) transformer = FluxTransformer2DModel.from_pretrained( "black-forest-labs/FLUX.1-dev", subfolder='transformer', torch_dtype=torch.bfloat16 ) vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae").to("cuda") pipe = FluxControlNetInpaintingPipeline.from_pretrained( "black-forest-labs/FLUX.1-dev", controlnet=controlnet, transformer=transformer, vae=vae, torch_dtype=torch.bfloat16 ) repo_name = "ByteDance/Hyper-SD" ckpt_name = "Hyper-FLUX.1-dev-8steps-lora.safetensors" pipe.load_lora_weights(hf_hub_download(repo_name, ckpt_name)) pipe.fuse_lora(lora_scale=0.125) pipe.transformer.to(torch.bfloat16) pipe.controlnet.to(torch.bfloat16) pipe.to("cuda") def can_expand(source_width, source_height, target_width, target_height, alignment): if alignment in ("Left", "Right") and source_width >= target_width: return False if alignment in ("Top", "Bottom") and source_height >= target_height: return False return True def prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom): target_size = (width, height) # Calculate the scaling factor to fit the image within the target size scale_factor = min(target_size[0] / image.width, target_size[1] / image.height) new_width = int(image.width * scale_factor) new_height = int(image.height * scale_factor) # Resize the source image to fit within target size source = image.resize((new_width, new_height), Image.LANCZOS) # Apply resize option using percentages if resize_option == "Full": resize_percentage = 100 elif resize_option == "50%": resize_percentage = 50 elif resize_option == "33%": resize_percentage = 33 elif resize_option == "25%": resize_percentage = 25 else: # Custom resize_percentage = custom_resize_percentage # Calculate new dimensions based on percentage resize_factor = resize_percentage / 100 new_width = int(source.width * resize_factor) new_height = int(source.height * resize_factor) # Ensure minimum size of 64 pixels new_width = max(new_width, 64) new_height = max(new_height, 64) # Resize the image source = source.resize((new_width, new_height), Image.LANCZOS) # Calculate the overlap in pixels based on the percentage overlap_x = int(new_width * (overlap_percentage / 100)) overlap_y = int(new_height * (overlap_percentage / 100)) # Ensure minimum overlap of 1 pixel overlap_x = max(overlap_x, 1) overlap_y = max(overlap_y, 1) # Calculate margins based on alignment if alignment == "Middle": margin_x = (target_size[0] - new_width) // 2 margin_y = (target_size[1] - new_height) // 2 elif alignment == "Left": margin_x = 0 margin_y = (target_size[1] - new_height) // 2 elif alignment == "Right": margin_x = target_size[0] - new_width margin_y = (target_size[1] - new_height) // 2 elif alignment == "Top": margin_x = (target_size[0] - new_width) // 2 margin_y = 0 elif alignment == "Bottom": margin_x = (target_size[0] - new_width) // 2 margin_y = target_size[1] - new_height # Adjust margins to eliminate gaps margin_x = max(0, min(margin_x, target_size[0] - new_width)) margin_y = max(0, min(margin_y, target_size[1] - new_height)) # Create a new background image and paste the resized source image background = Image.new('RGB', target_size, (255, 255, 255)) background.paste(source, (margin_x, margin_y)) # Create the mask mask = Image.new('L', target_size, 255) mask_draw = ImageDraw.Draw(mask) # Calculate overlap areas white_gaps_patch = 2 left_overlap = margin_x + overlap_x if overlap_left else margin_x + white_gaps_patch right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width - white_gaps_patch top_overlap = margin_y + overlap_y if overlap_top else margin_y + white_gaps_patch bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height - white_gaps_patch if alignment == "Left": left_overlap = margin_x + overlap_x if overlap_left else margin_x elif alignment == "Right": right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width elif alignment == "Top": top_overlap = margin_y + overlap_y if overlap_top else margin_y elif alignment == "Bottom": bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height # Draw the mask mask_draw.rectangle([ (left_overlap, top_overlap), (right_overlap, bottom_overlap) ], fill=0) return background, mask @spaces.GPU def inpaint(image, width, height, overlap_percentage, num_inference_steps, resize_option, custom_resize_percentage, prompt_input, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom, progress=gr.Progress(track_tqdm=True)): background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom) if not can_expand(background.width, background.height, width, height, alignment): alignment = "Middle" cnet_image = background.copy() cnet_image.paste(0, (0, 0), mask) final_prompt = f"{prompt_input} , high quality, 4k" generator = torch.Generator(device="cuda").manual_seed(42) result = pipe( prompt=final_prompt, height=height, width=width, control_image=cnet_image, control_mask=mask, num_inference_steps=num_inference_steps, generator=generator, controlnet_conditioning_scale=0.9, guidance_scale=3.5, negative_prompt="", true_guidance_scale=3.5, output_type="latent" ).images[0] pipe.to("cpu") vae.to("cuda") result = vae.decode(latent_image).sample result = result.convert("RGBA") cnet_image.paste(result, (0, 0), mask) return cnet_image, background def preview_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom): background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom) preview = background.copy().convert('RGBA') red_overlay = Image.new('RGBA', background.size, (255, 0, 0, 64)) red_mask = Image.new('RGBA', background.size, (0, 0, 0, 0)) red_mask.paste(red_overlay, (0, 0), mask) preview = Image.alpha_composite(preview, red_mask) return preview def clear_result(): return gr.update(value=None) def preload_presets(target_ratio, ui_width, ui_height): if target_ratio == "9:16": return 720, 1280, gr.update() elif target_ratio == "16:9": return 1280, 720, gr.update() elif target_ratio == "1:1": return 1024, 1024, gr.update() elif target_ratio == "Custom": return ui_width, ui_height, gr.update(open=True) def select_the_right_preset(user_width, user_height): if user_width == 720 and user_height == 1280: return "9:16" elif user_width == 1280 and user_height == 720: return "16:9" elif user_width == 1024 and user_height == 1024: return "1:1" else: return "Custom" def toggle_custom_resize_slider(resize_option): return gr.update(visible=(resize_option == "Custom")) def update_history(new_image, history): if history is None: history = [] history.insert(0, new_image) return history css = """ .gradio-container { width: 1200px !important; } """ title = """