Update app.py

app.py

@@ -1,95 +1,96 @@
-import spaces
-import gradio as gr
 import torch
+import torchvision.transforms as transforms
 from PIL import Image
-from diffusers import DiffusionPipeline
-import random
-
-# Initialize the base model and specific LoRA
-base_model = "black-forest-labs/FLUX.1-dev"
-pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16)
-pipe.to("cuda")
-
-lora_repo = "XLabs-AI/flux-RealismLora"
-trigger_word = ""  # Leave trigger_word blank if not used.
-pipe.load_lora_weights(lora_repo)
-
-MAX_SEED = 2**32-1
-
-@spaces.GPU(duration=80)
-def run_lora(prompt, cfg_scale, steps, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)):
-    # Set random seed for reproducibility
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator(device="cuda").manual_seed(seed)
-
-    # Update progress bar (0% saat mulai)
-    progress(0, "Starting image generation...")
-
-    # Generate image with progress updates
-    for i in range(1, steps + 1):
-        # Simulate the processing step (in a real scenario, you would integrate this with your image generation process)
-        if i % (steps // 10) == 0:  # Update every 10% of the steps
-            progress(i / steps * 100, f"Processing step {i} of {steps}...")
-
-    # Generate image using the pipeline
-    image = pipe(
-        prompt=f"{prompt} {trigger_word}",
-        num_inference_steps=steps,
-        guidance_scale=cfg_scale,
-        width=width,
-        height=height,
-        generator=generator,
-        joint_attention_kwargs={"scale": lora_scale},
-    ).images[0]
-
-    # Final update (100%)
-    progress(100, "Completed!")
-
-    yield image, seed
-
-# Example cached image and settings
-example_image_path = "example0.webp"  # Replace with the actual path to the example image
-example_prompt = """A Jelita Sukawati speaker is captured mid-speech. She has long, dark brown hair that cascades over her shoulders, framing her radiant, smiling face. Her Latina features are highlighted by warm, sun-kissed skin and bright, expressive eyes. She gestures with her left hand, displaying a delicate ring on her pinky finger, as she speaks passionately.
-The woman is wearing a colorful, patterned dress with a green lanyard featuring multiple badges and logos hanging around her neck. The lanyard prominently displays the "CagliostroLab" text.
-Behind her, there is a blurred background with a white banner containing logos and text, indicating a professional or conference setting. The overall scene captures the energy and vibrancy of her presentation."""
-example_cfg_scale = 3.2
-example_steps = 32
-example_width = 1152
-example_height = 896
-example_seed = 3981632454
-example_lora_scale = 0.85
-
-def load_example():
-    # Load example image from file
-    example_image = Image.open(example_image_path)
-    return example_prompt, example_cfg_scale, example_steps, False, example_seed, example_width, example_height, example_lora_scale, example_image
-
-with gr.Blocks() as app:
-    gr.Markdown("# Flux RealismLora Image Generator")
-    with gr.Row():
-        with gr.Column(scale=3):
-            prompt = gr.TextArea(label="Prompt", placeholder="Type a prompt", lines=5)
-            generate_button = gr.Button("Generate")
-            cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=example_cfg_scale)
-            steps = gr.Slider(label="Steps", minimum=1, maximum=100, step=1, value=example_steps)
-            width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=example_width)
-            height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=example_height)
-            randomize_seed = gr.Checkbox(False, label="Randomize seed")
-            seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=example_seed)
-            lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=1, step=0.01, value=example_lora_scale)
-        with gr.Column(scale=1):
-            result = gr.Image(label="Generated Image")
-            gr.Markdown("Generate images using RealismLora and a text prompt.\n[[non-commercial license, Flux.1 Dev](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md)]")
-
-    # Automatically load example data and image when the interface is launched
-    app.load(load_example, inputs=[], outputs=[prompt, cfg_scale, steps, randomize_seed, seed, width, height, lora_scale, result])
+import cv2
+import numpy as np
+import gradio as gr
+
+# Load MiDaS model
+midas = torch.hub.load("intel-isl/MiDaS", "DPT_Large")
+midas.eval()
+
+# Preprocessing function
+def preprocess_image(image):
+    transform = transforms.Compose([
+        transforms.Resize(384),
+        transforms.CenterCrop(384),
+        transforms.ToTensor(),
+        transforms.Normalize(
+            mean=[0.485, 0.456, 0.406],
+            std=[0.229, 0.224, 0.225],
+        ),
+    ])
+    return transform(image).unsqueeze(0)
+
+# Function to generate the displacement map
+def generate_displacement_map(image_a):
+    input_batch = preprocess_image(image_a)
+
+    with torch.no_grad():
+        depth_map = midas(input_batch)
+
+    depth_map = depth_map.squeeze().cpu().numpy()
+    depth_map = cv2.resize(depth_map, (image_a.width, image_a.height))
+
+    depth_map = (depth_map - depth_map.min()) / (depth_map.max() - depth_map.min())
+    displacement_map = depth_map * 30
+    return displacement_map
+
+# Function to warp and fit Image-B onto Image-A
+def fit_and_warp_design(image_a, image_b, design_bbox):
+    displacement_map = generate_displacement_map(image_a)
+
+    # Extract bounding box coordinates
+    top_left = (int(design_bbox[0]), int(design_bbox[1]))
+    bottom_right = (int(design_bbox[2]), int(design_bbox[3]))
+
+    # Resize the design to fit within the specified bounding box
+    design_width = bottom_right[0] - top_left[0]
+    design_height = bottom_right[1] - top_left[1]
+    image_b = image_b.resize((design_width, design_height))
     
-    generate_button.click(
-        run_lora,
-        inputs=[prompt, cfg_scale, steps, randomize_seed, seed, width, height, lora_scale],
-        outputs=[result, seed]
-    )
-
-app.queue()
-app.launch()
+    # Create a blank canvas with the same size as Image-A
+    canvas = Image.new('RGBA', (displacement_map.shape[1], displacement_map.shape[0]), (0, 0, 0, 0))
+    canvas.paste(image_b, top_left, image_b)
+    canvas_np = np.array(canvas)
+
+    h, w = displacement_map.shape
+    y_indices, x_indices = np.indices((h, w), dtype=np.float32)
+    x_displacement = (x_indices + displacement_map).astype(np.float32)
+    y_displacement = (y_indices + displacement_map).astype(np.float32)
+
+    x_displacement = np.clip(x_displacement, 0, w - 1)
+    y_displacement = np.clip(y_displacement, 0, h - 1)
+
+    warped_canvas = cv2.remap(canvas_np, x_displacement, y_displacement, cv2.INTER_LINEAR, borderMode=cv2.BORDER_TRANSPARENT)
+
+    image_a_rgba = image_a.convert("RGBA")
+    image_a_np = np.array(image_a_rgba)
+
+    non_transparent_pixels = warped_canvas[..., 3] > 0
+    image_a_np[non_transparent_pixels] = warped_canvas[non_transparent_pixels]
+
+    final_image = Image.fromarray(image_a_np)
+    return final_image
+
+# Gradio interface function
+def process_images(image_a, image_b, design_bbox):
+    result = fit_and_warp_design(image_a, image_b, design_bbox)
+    return result
+
+# Gradio UI components
+image_input_a = gr.inputs.Image(label="Upload Clothing Image", type="pil")
+image_input_b = gr.inputs.Image(label="Upload Design Image", type="pil")
+design_bbox_input = gr.inputs.Image(tool="select", label="Adjust Design Position and Size")
+
+# Define the Gradio interface
+iface = gr.Interface(
+    fn=process_images,
+    inputs=[image_input_a, design_bbox_input],
+    outputs="image",
+    title="Clothing Design Fitting with Drag-and-Drop",
+    description="Upload a clothing image and a design image. Drag and resize the design onto the clothing using the cursor.",
+)
+
+# Launch the interface
+iface.launch()