import os import gradio as gr import numpy as np import spaces import torch import random from PIL import Image from typing import Iterable # --- Gradio Theme --- from gradio.themes import Soft from gradio.themes.utils import colors, fonts, sizes colors.steel_blue = colors.Color( name="steel_blue", c50="#EBF3F8", c100="#D3E5F0", c200="#A8CCE1", c300="#7DB3D2", c400="#529AC3", c500="#4682B4", c600="#3E72A0", c700="#36638C", c800="#2E5378", c900="#264364", c950="#1E3450", ) class SteelBlueTheme(Soft): def __init__( self, *, primary_hue: colors.Color | str = colors.gray, secondary_hue: colors.Color | str = colors.steel_blue, neutral_hue: colors.Color | str = colors.slate, text_size: sizes.Size | str = sizes.text_lg, font: fonts.Font | str | Iterable[fonts.Font | str] = ( fonts.GoogleFont("Outfit"), "Arial", "sans-serif", ), font_mono: fonts.Font | str | Iterable[fonts.Font | str] = ( fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace", ), ): super().__init__( primary_hue=primary_hue, secondary_hue=secondary_hue, neutral_hue=neutral_hue, text_size=text_size, font=font, font_mono=font_mono, ) super().set( background_fill_primary="*primary_50", background_fill_primary_dark="*primary_900", body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)", body_background_fill_dark="linear-gradient(135deg, *primary_900, *primary_800)", button_primary_text_color="white", button_primary_text_color_hover="white", button_primary_background_fill="linear-gradient(90deg, *secondary_500, *secondary_600)", button_primary_background_fill_hover="linear-gradient(90deg, *secondary_600, *secondary_700)", slider_color="*secondary_500", slider_color_dark="*secondary_600", block_title_text_weight="600", block_border_width="3px", block_shadow="*shadow_drop_lg", ) steel_blue_theme = SteelBlueTheme() # --- Model Loading --- from diffusers import FlowMatchEulerDiscreteScheduler # from optimization import optimize_pipeline_ # Assuming this is a custom file from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3 dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" pipe = QwenImageEditPlusPipeline.from_pretrained( "Qwen/Qwen-Image-Edit-2509", transformer=QwenImageTransformer2DModel.from_pretrained( "linoyts/Qwen-Image-Edit-Rapid-AIO", subfolder='transformer', torch_dtype=dtype, device_map='cuda' ), torch_dtype=dtype ).to(device) # Load all LoRA adapters pipe.load_lora_weights("autoweeb/Qwen-Image-Edit-2509-Photo-to-Anime", weight_name="Qwen-Image-Edit-2509-Photo-to-Anime_000001000.safetensors", adapter_name="anime") pipe.load_lora_weights("dx8152/Qwen-Edit-2509-Multiple-angles", weight_name="镜头转换.safetensors", adapter_name="multiple-angles") pipe.load_lora_weights("dx8152/Qwen-Image-Edit-2509-Light_restoration", weight_name="移除光影.safetensors", adapter_name="light-restoration") pipe.load_lora_weights("dx8152/Qwen-Image-Edit-2509-Relight", weight_name="Qwen-Edit-Relight.safetensors", adapter_name="relight") pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3()) MAX_SEED = np.iinfo(np.int32).max # --- Helper Function for Aspect Ratio --- def update_dimensions_on_upload(image): if image is None: return 1024, 1024 original_width, original_height = image.size # Cap max dimension to 1024 while preserving aspect ratio if original_width > original_height: new_width = 1024 new_height = int(1024 * original_height / original_width) else: new_height = 1024 new_width = int(1024 * original_width / original_height) # Ensure dimensions are multiples of 8 for model compatibility new_width = (new_width // 8) * 8 new_height = (new_height // 8) * 8 return new_width, new_height # --- Main Inference Function --- @spaces.GPU def infer( input_image, prompt, lora_adapter, seed, randomize_seed, guidance_scale, steps, width, height, progress=gr.Progress(track_tqdm=True) ): if input_image is None: raise gr.Error("Please upload an image to edit.") # Dynamically set the adapter if lora_adapter == "Photo-to-Anime": pipe.set_adapters(["anime"], adapter_weights=[1.0]) elif lora_adapter == "Multiple-Angles": pipe.set_adapters(["multiple-angles"], adapter_weights=[1.0]) elif lora_adapter == "Light-Restoration": pipe.set_adapters(["light-restoration"], adapter_weights=[1.0]) elif lora_adapter == "Relight": pipe.set_adapters(["relight"], adapter_weights=[1.0]) if randomize_seed: seed = random.randint(0, MAX_SEED) generator = torch.Generator(device=device).manual_seed(seed) negative_prompt = "worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry" result = pipe( image=input_image.convert("RGB"), prompt=prompt, negative_prompt=negative_prompt, height=height, width=width, num_inference_steps=steps, generator=generator, true_cfg_scale=guidance_scale, num_images_per_prompt=1, ).images[0] # *** FIX: Changed function to return only 2 values to match the button's expectation *** return result, seed # --- Wrapper for Examples --- @spaces.GPU def infer_example(input_image_path, prompt, lora_adapter): # *** FIX: Fully implemented this function to handle examples correctly *** input_pil = Image.open(input_image_path).convert("RGB") # Calculate aspect ratio for the example image width, height = update_dimensions_on_upload(input_pil) # Set reasonable default values for example inference guidance_scale = 1.0 steps = 4 # Call the main infer function result, seed = infer(input_pil, prompt, lora_adapter, 0, True, guidance_scale, steps, width, height) return result, seed # --- UI Layout --- css=""" #col-container { margin: 0 auto; max-width: 960px; } #main-title h1 {font-size: 2.1em !important;} """ with gr.Blocks(css=css, theme=steel_blue_theme) as demo: with gr.Column(elem_id="col-container"): gr.Markdown("# **Qwen-Image-Edit-2509-LoRAs-Fast**", elem_id="main-title") gr.Markdown("Perform diverse image edits using specialized LoRA adapters for the Qwen-Image-Edit model.") with gr.Row(equal_height=True): with gr.Column(): input_image = gr.Image(label="Upload Image", type="pil", height=400) lora_adapter = gr.Dropdown( label="Choose Editing Style", choices=["Photo-to-Anime", "Multiple-Angles", "Light-Restoration", "Relight"], value="Photo-to-Anime" ) prompt = gr.Text( label="Edit Prompt", show_label=True, placeholder="e.g., transform into anime", ) run_button = gr.Button("Run", variant="primary") with gr.Accordion("⚙️ Advanced Settings", open=False): seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0) randomize_seed = gr.Checkbox(label="Randomize Seed", value=True) guidance_scale = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=10.0, step=0.1, value=1.0) steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=4) # Hidden sliders to hold image dimensions height = gr.Slider(label="Height", minimum=256, maximum=1024, step=8, value=1024, visible=False) width = gr.Slider(label="Width", minimum=256, maximum=1024, step=8, value=1024, visible=False) with gr.Column(): output_image = gr.Image(label="Output Image", interactive=False, format="png", height=400) gr.Examples( examples=[ ["examples/1.jpg", "Transform into anime.", "Photo-to-Anime"], ["examples/4.jpg", "Remove shadows and relight the image using soft lighting.", "Light-Restoration"], ["examples/5.jpg", "Relight the image using soft, diffused lighting that simulates sunlight filtering through curtains.", "Relight"], ["examples/2.jpeg", "Move the camera left.", "Multiple-Angles"], ["examples/2.jpeg", "Move the camera right.", "Multiple-Angles"], ["examples/2.jpeg", "Move the camera down.", "Multiple-Angles"], ["examples/2.jpeg", "Rotate the camera 45 degrees to the left.", "Multiple-Angles"], ["examples/3.jpg", "Rotate the camera 45 degrees to the right.", "Multiple-Angles"], ["examples/3.jpg", "Switch the camera to a top-down view.", "Multiple-Angles"], ["examples/3.jpg", "Switch the camera to a wide-angle lens.", "Multiple-Angles"], ["examples/3.jpg", "Switch the camera to a close-up lens.", "Multiple-Angles"], ], inputs=[input_image, prompt, lora_adapter], outputs=[output_image, seed], fn=infer_example, cache_examples="lazy", # Changed to lazy for better performance label="Examples" ) # --- Event Handlers --- run_button.click( fn=infer, inputs=[input_image, prompt, lora_adapter, seed, randomize_seed, guidance_scale, steps, width, height], outputs=[output_image, seed] ) input_image.upload( fn=update_dimensions_on_upload, inputs=[input_image], outputs=[width, height] ) demo.launch()