Update app.py

app.py

@@ -3,51 +3,77 @@ import spaces
 import torch
 from diffusers import AutoencoderKL, TCDScheduler
 from diffusers.models.model_loading_utils import load_state_dict
-# Remove ImageSlider import
-# from gradio_imageslider import ImageSlider
+# Removed ImageSlider import
 from huggingface_hub import hf_hub_download
 
-from controlnet_union import ControlNetModel_Union
-from pipeline_fill_sd_xl import StableDiffusionXLFillPipeline
+# Ensure these custom modules are accessible in the environment
+# If running locally, they should be in the same directory or installed
+try:
+    from controlnet_union import ControlNetModel_Union
+    from pipeline_fill_sd_xl import StableDiffusionXLFillPipeline
+except ImportError as e:
+    print(f"Error importing custom modules: {e}")
+    print("Please ensure 'controlnet_union.py' and 'pipeline_fill_sd_xl.py' are in the working directory or installed.")
+    # Optionally, try installing if running in a suitable environment
+    # import os
+    # os.system("pip install git+https://github.com/UNION-AI-Research/FILL-Context-Aware-Outpainting.git") # Or wherever the package is hosted
+    # Re-try import might be needed depending on environment setup
+    exit()
+
 
 from PIL import Image, ImageDraw
 import numpy as np
+import os # For checking example files
+
+# --- Model Loading ---
+# Use environment variable for model cache if needed
+# HUGGINGFACE_HUB_CACHE = os.environ.get("HUGGINGFACE_HUB_CACHE", None)
+
+try:
+    config_file = hf_hub_download(
+        "xinsir/controlnet-union-sdxl-1.0",
+        filename="config_promax.json",
+        # cache_dir=HUGGINGFACE_HUB_CACHE
+    )
+
+    config = ControlNetModel_Union.load_config(config_file)
+    controlnet_model = ControlNetModel_Union.from_config(config)
+    model_file = hf_hub_download(
+        "xinsir/controlnet-union-sdxl-1.0",
+        filename="diffusion_pytorch_model_promax.safetensors",
+        # cache_dir=HUGGINGFACE_HUB_CACHE
+    )
 
-# --- Model Loading (Unchanged) ---
-config_file = hf_hub_download(
-    "xinsir/controlnet-union-sdxl-1.0",
-    filename="config_promax.json",
-)
-
-config = ControlNetModel_Union.load_config(config_file)
-controlnet_model = ControlNetModel_Union.from_config(config)
-model_file = hf_hub_download(
-    "xinsir/controlnet-union-sdxl-1.0",
-    filename="diffusion_pytorch_model_promax.safetensors",
-)
-
-sstate_dict = load_state_dict(model_file)
-model, _, _, _, _ = ControlNetModel_Union._load_pretrained_model(
-    controlnet_model, sstate_dict, model_file, "xinsir/controlnet-union-sdxl-1.0"
-)
-model.to(device="cuda", dtype=torch.float16)
-#----------------------
-
-vae = AutoencoderKL.from_pretrained(
-    "madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16
-).to("cuda")
-
-pipe = StableDiffusionXLFillPipeline.from_pretrained(
-    "SG161222/RealVisXL_V5.0_Lightning",
-    torch_dtype=torch.float16,
-    vae=vae,
-    controlnet=model,
-    variant="fp16",
-).to("cuda")
-
-pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
-
-# --- Helper Functions (Mostly Unchanged) ---
+    sstate_dict = load_state_dict(model_file)
+    model, _, _, _, _ = ControlNetModel_Union._load_pretrained_model(
+        controlnet_model, sstate_dict, model_file, "xinsir/controlnet-union-sdxl-1.0"
+    )
+    model.to(device="cuda", dtype=torch.float16)
+    print("ControlNet loaded successfully.")
+
+    vae = AutoencoderKL.from_pretrained(
+        "madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16, # cache_dir=HUGGINGFACE_HUB_CACHE
+    ).to("cuda")
+    print("VAE loaded successfully.")
+
+    pipe = StableDiffusionXLFillPipeline.from_pretrained(
+        "SG161222/RealVisXL_V5.0_Lightning",
+        torch_dtype=torch.float16,
+        vae=vae,
+        controlnet=model,
+        variant="fp16",
+        # cache_dir=HUGGINGFACE_HUB_CACHE
+    ).to("cuda")
+    print("Pipeline loaded successfully.")
+
+    pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
+    print("Scheduler configured.")
+
+except Exception as e:
+    print(f"Error during model loading: {e}")
+    raise e
+
+# --- Helper Functions ---
 def can_expand(source_width, source_height, target_width, target_height, alignment):
     """Checks if the image can be expanded based on the alignment."""
     if alignment in ("Left", "Right") and source_width >= target_width:
@@ -57,211 +83,305 @@ def can_expand(source_width, source_height, target_width, target_height, alignme
     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 # Pixels to leave unmasked at edges if overlap is disabled for that edge
-
-    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
-
-    # Adjust overlap boundaries based on alignment when specific overlap directions are *disabled*
-    # This prevents unmasking the absolute edge of the canvas in alignment modes
-    if alignment == "Left":
-        left_overlap = margin_x + overlap_x if overlap_left else margin_x # Keep edge masked if alignment is left
-    elif alignment == "Right":
-        right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width # Keep edge masked
-    elif alignment == "Top":
-        top_overlap = margin_y + overlap_y if overlap_top else margin_y # Keep edge masked
-    elif alignment == "Bottom":
-        bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height # Keep edge masked
-
-    # Ensure coordinates are within bounds
-    left_overlap = max(0, left_overlap)
-    top_overlap = max(0, top_overlap)
-    right_overlap = min(target_size[0], right_overlap)
-    bottom_overlap = min(target_size[1], bottom_overlap)
-
-    # Draw the mask (black rectangle for the area to keep)
-    if right_overlap > left_overlap and bottom_overlap > top_overlap:
-        mask_draw.rectangle([
-            (left_overlap, top_overlap),
-            (right_overlap, bottom_overlap)
-        ], fill=0) # 0 means keep this area (not masked for inpainting)
-
-    # Invert the mask: White areas (255) will be inpainted. Black (0) is kept.
-    mask = Image.fromarray(255 - np.array(mask))
-
-    return background, mask
+    if image is None:
+        raise gr.Error("Input image not provided.")
+    try:
+        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
+        elif resize_option == "Custom":
+            resize_percentage = custom_resize_percentage
+        else:
+             raise ValueError(f"Invalid resize option: {resize_option}")
+
+
+        # 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)
+
+        # Ensure dimensions fit within target (can happen if original image is tiny and resize % is large)
+        new_width = min(new_width, target_size[0])
+        new_height = min(new_height, target_size[1])
+
+        # 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 if overlap is enabled, otherwise 0
+        overlap_x = max(overlap_x, 1) if overlap_left or overlap_right else 0
+        overlap_y = max(overlap_y, 1) if overlap_top or overlap_bottom else 0
+
+        # 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
+        else:
+             raise ValueError(f"Invalid alignment: {alignment}")
+
+
+        # Adjust margins to ensure image is fully within bounds (should be redundant with min check above)
+        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)) # White background
+        background.paste(source, (margin_x, margin_y))
+
+        # Create the mask (initially all black - meaning keep everything)
+        mask_np = np.zeros(target_size[::-1], dtype=np.uint8) # Use numpy for easier slicing [::-1] for (height, width)
+
+        # Calculate the coordinates of the *source image* area within the target canvas
+        source_left = margin_x
+        source_top = margin_y
+        source_right = margin_x + new_width
+        source_bottom = margin_y + new_height
+
+        # Calculate the coordinates of the *unmasked* area (area to keep from source)
+        unmasked_left = source_left + overlap_x if overlap_left else source_left
+        unmasked_top = source_top + overlap_y if overlap_top else source_top
+        unmasked_right = source_right - overlap_x if overlap_right else source_right
+        unmasked_bottom = source_bottom - overlap_y if overlap_bottom else source_bottom
+
+        # Special handling for edge alignments to ensure the edge itself is kept if overlap disabled
+        if alignment == "Left" and not overlap_left:
+            unmasked_left = source_left
+        if alignment == "Right" and not overlap_right:
+            unmasked_right = source_right
+        if alignment == "Top" and not overlap_top:
+            unmasked_top = source_top
+        if alignment == "Bottom" and not overlap_bottom:
+            unmasked_bottom = source_bottom
+
+        # Ensure coordinates are valid and clipped to the source image area within the canvas
+        unmasked_left = max(source_left, min(unmasked_left, source_right))
+        unmasked_top = max(source_top, min(unmasked_top, source_bottom))
+        unmasked_right = max(source_left, min(unmasked_right, source_right))
+        unmasked_bottom = max(source_top, min(unmasked_bottom, source_bottom))
+
+        # Create the final mask: White (255) = Area to inpaint/outpaint, Black (0) = Area to keep
+        final_mask_np = np.ones(target_size[::-1], dtype=np.uint8) * 255 # Start with all white (change everything)
+        if unmasked_right > unmasked_left and unmasked_bottom > unmasked_top:
+             # Set the area to keep (calculated unmasked rectangle) to black (0)
+             final_mask_np[unmasked_top:unmasked_bottom, unmasked_left:unmasked_right] = 0
+
+        mask = Image.fromarray(final_mask_np)
+
+        return background, mask
+    except Exception as e:
+        print(f"Error in prepare_image_and_mask: {e}")
+        raise gr.Error(f"Failed to prepare image and mask: {e}")
+
 
 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)
-
-    # Create a preview image showing the mask
-    preview = background.copy().convert('RGBA')
-
-    # Create a semi-transparent red overlay for the masked (inpainting) area
-    red_overlay = Image.new('RGBA', background.size, (255, 0, 0, 100)) # 100 alpha (~40% opacity)
-
-    # The mask is now white (255) where inpainting happens. Use this directly.
-    preview.paste(red_overlay, (0, 0), mask)
-
-    return preview
-
-@spaces.GPU(duration=24)
-def infer(image, width, height, overlap_percentage, num_inference_steps, resize_option, custom_resize_percentage, prompt_input, 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)
-
-    # Ensure alignment allows expansion, default to Middle if not
-    source_w, source_h = background.size # Use background size after initial resize/placement
-    target_w, target_h = width, height
-    if alignment in ("Left", "Right") and source_w >= target_w:
-        print(f"Warning: Source width ({source_w}) >= target width ({target_w}) with {alignment} alignment. Forcing Middle alignment.")
-        alignment = "Middle"
-        # Re-prepare mask/background with corrected alignment if needed (optional, depends if prepare func uses alignment early)
-        # background, mask = prepare_image_and_mask(...) # If needed
-    if alignment in ("Top", "Bottom") and source_h >= target_h:
-        print(f"Warning: Source height ({source_h}) >= target height ({target_h}) with {alignment} alignment. Forcing Middle alignment.")
-        alignment = "Middle"
-        # Re-prepare mask/background with corrected alignment if needed
-        # background, mask = prepare_image_and_mask(...) # If needed
-
-    # Image for ControlNet input (masked original content)
-    # The pipeline expects the original image content in the non-masked area
-    cnet_image = background.copy()
-    # The pipeline's `image` argument is the *initial* content for the *masked* area (often noise, but here we provide the background)
-    # The `mask_image` tells the pipeline *where* to perform the inpainting/outpainting.
-    # The controlnet `image` needs the original content visible in the non-masked area.
-    # ControlNet Union seems to work well by just passing the background with the source image pasted.
-
-    final_prompt = f"{prompt_input} , high quality, 4k" if prompt_input else "high quality, 4k"
-
-    (
-        prompt_embeds,
-        negative_prompt_embeds,
-        pooled_prompt_embeds,
-        negative_pooled_prompt_embeds,
-    ) = pipe.encode_prompt(final_prompt, "cuda", True)
-
-    # The pipeline call
-    # Note: The pipeline expects `image` (initial state for masked area) and `mask_image`
-    # The `control_image` is implicitly handled by the ControlNet attached to the pipeline
-    output_image = pipe(
-        prompt_embeds=prompt_embeds,
-        negative_prompt_embeds=negative_prompt_embeds,
-        pooled_prompt_embeds=pooled_prompt_embeds,
-        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
-        image=background, # Provide the initial canvas state
-        mask_image=mask,   # Provide the mask (white is area to change)
-        control_image=cnet_image, # Pass the control image explicitly if needed by pipeline logic
-        num_inference_steps=num_inference_steps,
-        output_type="pil" # Ensure PIL output
-    ).images[0]
-
-    # The pipeline should have already handled the compositing based on the mask
-    # If not, uncomment the paste operation below:
-    # final_image = background.copy().convert("RGBA") # Start with original background
-    # output_image = output_image.convert("RGBA")
-    # mask_rgba = mask.convert('L').point(lambda p: 255 if p > 128 else 0) # Ensure mask is binary 0/255
-    # final_image.paste(output_image, (0, 0), mask_rgba) # Paste generated content using the mask
-
-    # Return the single final image
-    return output_image
-
-
-def clear_result():
-    """Clears the result Image component."""
-    return gr.update(value=None)
-
-# --- UI Helper Functions (Unchanged) ---
+    if image is None:
+        return None # Or return a placeholder image/message
+    try:
+        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)
+
+        # Create a preview image showing the mask
+        preview = background.copy().convert('RGBA')
+
+        # Create a semi-transparent red overlay for the masked (inpainting/outpainting) area
+        red_overlay = Image.new('RGBA', background.size, (255, 0, 0, 100)) # 100 alpha (~40% opacity)
+
+        # The mask is white (255) where outpainting happens. Use this directly.
+        preview.paste(red_overlay, (0, 0), mask) # Paste red where mask is white
+
+        return preview
+    except Exception as e:
+        print(f"Error during preview generation: {e}")
+        # Return the original background or an error placeholder
+        if 'background' in locals():
+             return background.convert('RGBA')
+        else:
+             return Image.new('RGBA', (width, height), (200, 200, 200, 255)) # Grey placeholder
+
+
+@spaces.GPU(duration=60) # Adjusted duration slightly
+def infer(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)):
+    if image is None:
+         raise gr.Error("Please provide an input image.")
+
+    try:
+        # --- Preparation ---
+        progress(0.1, desc="Preparing image and mask...")
+        original_alignment = alignment
+        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)
+
+        # --- Alignment Check & Correction ---
+        # Get dimensions *after* initial placement and resize
+        pasted_source_img_width = int(image.width * min(width / image.width, height / image.height) * (custom_resize_percentage if resize_option=='Custom' else {'Full':100, '50%':50, '33%':33, '25%':25}[resize_option])/100)
+        pasted_source_img_height = int(image.height * min(width / image.width, height / image.height) * (custom_resize_percentage if resize_option=='Custom' else {'Full':100, '50%':50, '33%':33, '25%':25}[resize_option])/100)
+        pasted_source_img_width = max(64, min(pasted_source_img_width, width))
+        pasted_source_img_height = max(64, min(pasted_source_img_height, height))
+
+        needs_reprepare = False
+        if alignment in ("Left", "Right") and pasted_source_img_width >= width:
+            print(f"Warning: Source width ({pasted_source_img_width}) >= target width ({width}) with {alignment} alignment. Forcing Middle alignment.")
+            alignment = "Middle"
+            needs_reprepare = True
+        if alignment in ("Top", "Bottom") and pasted_source_img_height >= height:
+            print(f"Warning: Source height ({pasted_source_img_height}) >= target height ({height}) with {alignment} alignment. Forcing Middle alignment.")
+            alignment = "Middle"
+            needs_reprepare = True
+
+        if needs_reprepare and alignment != original_alignment:
+            print("Re-preparing mask due to alignment change.")
+            progress(0.15, desc="Re-preparing mask for Middle alignment...")
+            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)
+
+        # ControlNet expects the image with the *original* content visible in the non-masked area
+        cnet_image = background.copy()
+        # In some ControlNet inpainting setups, you might mask the control image too,
+        # but Union ControlNet Fill often works well with the unmasked source pasted onto the background.
+        # cnet_image.paste(0, mask=ImageOps.invert(mask)) # Optional: Black out masked area in CNet image
+
+        # --- Prompt Encoding ---
+        progress(0.2, desc="Encoding prompt...")
+        final_prompt = f"{prompt_input}, high quality, 4k" if prompt_input else "high quality, 4k" # Add default tags if no prompt
+        negative_prompt = "low quality, blurry, noisy, text, words, letters, watermark, signature, username, artist name, deformed, distorted, disfigured, bad anatomy, extra limbs, missing limbs"
+
+
+        # Note: TCD/Lightning pipelines often work better *without* explicit negative prompts encoded
+        # Try encoding only the positive prompt first
+        (
+            prompt_embeds,
+            _, # negative_prompt_embeds (set to None or handle differently for TCD)
+            pooled_prompt_embeds,
+            _, # negative_pooled_prompt_embeds
+        ) = pipe.encode_prompt(final_prompt, "cuda", False) # do_classifier_free_guidance=False for TCD
+
+
+        # --- Inference ---
+        progress(0.3, desc="Starting diffusion process...")
+        print(f"Running inference with {num_inference_steps} steps...")
+        pipeline_output = pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=None, # Pass None for TCD/Lightning
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=None, # Pass None for TCD/Lightning
+            image=background,           # Initial state for masked area (background with source)
+            mask_image=mask,            # Mask (white = change)
+            control_image=cnet_image,   # ControlNet input
+            num_inference_steps=num_inference_steps,
+            guidance_scale=0.0,         # Crucial for TCD/Lightning
+            controlnet_conditioning_scale=0.8, # Default for FILL pipeline, adjust if needed
+            output_type="pil"           # Ensure PIL output
+            # Add tqdm=True if supported by the custom pipeline and using gr.Progress without track_tqdm
+        )
+
+        # --- Process Output ---
+        progress(0.9, desc="Processing results...")
+        # Check if the pipeline returned a standard output object or a generator
+        output_image = None
+        if hasattr(pipeline_output, 'images'): # Standard diffusers output
+            print("Pipeline returned a standard output object.")
+            if len(pipeline_output.images) > 0:
+                output_image = pipeline_output.images[0]
+            else:
+                 raise ValueError("Pipeline output contained no images.")
+        # Check if it's iterable (generator) - less likely with direct call and output_type='pil' but good practice
+        elif hasattr(pipeline_output, '__iter__') and not isinstance(pipeline_output, dict):
+            print("Pipeline returned a generator, iterating to get the final image.")
+            last_item = None
+            for item in pipeline_output:
+                last_item = item
+             # Try to extract image from the last yielded item (structure can vary)
+            if isinstance(last_item, tuple) and len(last_item) > 0 and isinstance(last_item[0], Image.Image):
+                output_image = last_item[0]
+            elif isinstance(last_item, dict) and 'images' in last_item and len(last_item['images']) > 0:
+                output_image = last_item['images'][0]
+            elif isinstance(last_item, Image.Image):
+                output_image = last_item
+            elif hasattr(last_item, 'images') and len(last_item.images) > 0: # Handle case where object yielded early
+                 output_image = last_item.images[0]
+
+            if output_image is None:
+                raise ValueError("Pipeline generator did not yield a valid final image structure.")
+        else:
+            raise TypeError(f"Unexpected pipeline output type: {type(pipeline_output)}. Cannot extract image.")
+
+        print("Inference complete.")
+        progress(1.0, desc="Done!")
+        return output_image
+
+    except Exception as e:
+        print(f"Error during inference: {e}")
+        import traceback
+        traceback.print_exc() # Print full traceback to console/logs
+        raise gr.Error(f"Inference failed: {e}")
+
+
+def clear_result(*args):
+    """Clears the result Image and related components."""
+    updates = {
+        result: gr.update(value=None),
+        use_as_input_button: gr.update(visible=False),
+    }
+    # If preview image is passed as an arg, clear it too
+    if len(args) > 0 and isinstance(args[0], gr.Image):
+         updates[args[0]] = gr.update(value=None) # Assuming preview_image is the first optional arg
+    return updates
+
+
+# --- UI Helper Functions ---
 def preload_presets(target_ratio, ui_width, ui_height):
     """Updates the width and height sliders based on the selected aspect ratio."""
+    settings_update = gr.update() # Default: no change to accordion state
     if target_ratio == "9:16":
         changed_width = 720
         changed_height = 1280
-        return changed_width, changed_height, gr.update() # Close accordion
     elif target_ratio == "16:9":
         changed_width = 1280
         changed_height = 720
-        return changed_width, changed_height, gr.update() # Close accordion
     elif target_ratio == "1:1":
         changed_width = 1024
         changed_height = 1024
-        return changed_width, changed_height, gr.update() # Close accordion
     elif target_ratio == "Custom":
-        # Don't change sliders, just open accordion
-        return ui_width, ui_height, gr.update(open=True)
+        changed_width = ui_width   # Keep current slider values
+        changed_height = ui_height
+        settings_update = gr.update(open=True) # Open accordion for custom
+    else: # Should not happen
+         changed_width = ui_width
+         changed_height = ui_height
+
+    return changed_width, changed_height, settings_update
 
 def select_the_right_preset(user_width, user_height):
     """Updates the radio button based on the current slider values."""
@@ -280,172 +400,254 @@ def toggle_custom_resize_slider(resize_option):
 
 def update_history(new_image, history):
     """Updates the history gallery with the new image."""
+    if not isinstance(new_image, Image.Image): # Don't add if generation failed (None)
+         return history or [] # Return current or empty list
+
     if history is None:
         history = []
-    # Ensure new_image is a PIL Image before adding
-    if isinstance(new_image, Image.Image):
-        history.insert(0, new_image)
+    history.insert(0, new_image)
+    # Limit history size (optional)
+    max_history = 12
+    if len(history) > max_history:
+        history = history[:max_history]
     return history
 
 # --- Gradio UI Definition ---
 css = """
 .gradio-container {
-    width: 1200px !important;
+    max-width: 1200px !important; /* Use max-width for responsiveness */
     margin: auto !important; /* Center the container */
+    padding: 10px; /* Add some padding */
 }
-h1 { text-align: center; }
-footer { visibility: hidden; }
+h1 { text-align: center; margin-bottom: 15px;}
+footer { display: none !important; /* More reliable way to hide footer */ }
+
 /* Ensure result image takes reasonable space */
 #result-image img {
     max-height: 768px; /* Adjust max height as needed */
     object-fit: contain;
-    width: auto;
+    width: 100%; /* Allow image to use column width */
     height: auto;
+    display: block; /* Prevent extra space below image */
+    margin: auto; /* Center image within its container */
+}
+#input-image img {
+     max-height: 400px;
+     object-fit: contain;
+     width: 100%;
+     height: auto;
+     display: block;
+     margin: auto;
+}
+#preview-image img {
+     max-height: 250px; /* Smaller preview */
+     object-fit: contain;
+     width: 100%;
+     height: auto;
+     display: block;
+     margin: auto;
 }
+
 #history-gallery .thumbnail-item { /* Style history items */
     height: 100px !important;
+    overflow: hidden; /* Hide overflow */
 }
 #history-gallery .gallery {
     grid-template-rows: repeat(auto-fill, 100px) !important;
+    gap: 4px !important; /* Add small gap */
+}
+#history-gallery .thumbnail-item img {
+    object-fit: contain !important; /* Ensure history previews fit */
+    height: 100%;
+    width: 100%;
 }
 
+/* Make Checkboxes smaller and closer */
+.gradio-checkboxgroup .wrap {
+    gap: 0.5rem 1rem !important; /* Adjust spacing */
+}
+.gradio-checkbox label span {
+    font-size: 0.9em; /* Slightly smaller label text */
+}
+.gradio-checkbox input {
+    transform: scale(0.9); /* Slightly smaller checkbox */
+}
+
+/* Style Accordion */
+.gradio-accordion .label-wrap { /* Target the label wrapper */
+    border: 1px solid #e0e0e0;
+    border-radius: 5px;
+    padding: 8px 12px;
+    background-color: #f9f9f9;
+}
 """
 
-title = """<h1 align="center">Diffusers Image Outpaint Lightning</h1>"""
+title = """<h1 align="center">🖼️ Diffusers Image Outpaint Lightning ⚡</h1>"""
 
-with gr.Blocks(css=css) as demo:
-    with gr.Column():
-        gr.HTML(title)
+# --- Example Files Handling ---
+# Create examples directory if it doesn't exist
+if not os.path.exists("./examples"):
+    os.makedirs("./examples")
 
-        with gr.Row():
-            with gr.Column(scale=1): # Left column for inputs
-                input_image = gr.Image(
-                    type="pil",
-                    label="Input Image",
-                    height=400 # Give input image reasonable height
+# Check for example images and provide defaults or placeholders if missing
+example_files = {
+    "ex1": "./examples/example_1.webp",
+    "ex2": "./examples/example_2.jpg",
+    "ex3": "./examples/example_3.jpg"
+}
+default_image_path = None # Will be set to the first available example
+
+# You might want to download example images if they don't exist
+# from huggingface_hub import hf_hub_download
+# def download_example(repo_id, filename, local_path):
+#     if not os.path.exists(local_path):
+#         try:
+#             hf_hub_download(repo_id=repo_id, filename=filename, local_dir="./examples", local_dir_use_symlinks=False)
+#             print(f"Downloaded {filename}")
+#         except Exception as e:
+#             print(f"Failed to download example {filename}: {e}")
+#             return False # Indicate failure
+#     return os.path.exists(local_path)
+
+# Example: download_example("path/to/your/example-repo", "example_1.webp", example_files["ex1"])
+# For now, we just check existence
+
+examples_available = {key: os.path.exists(path) for key, path in example_files.items()}
+
+example_list = []
+if examples_available["ex1"]:
+    example_list.append([example_files["ex1"], "A wide landscape view of the mountains", 1280, 720, "Middle"])
+    if default_image_path is None: default_image_path = example_files["ex1"]
+if examples_available["ex2"]:
+     example_list.append([example_files["ex2"], "Full body shot of the astronaut on the moon", 720, 1280, "Middle"])
+     if default_image_path is None: default_image_path = example_files["ex2"]
+if examples_available["ex3"]:
+     example_list.append([example_files["ex3"], "Expanding the sky and ground around the subject", 1024, 1024, "Middle"])
+     example_list.append([example_files["ex3"], "Expanding downwards from the subject", 1024, 1024, "Top"])
+     example_list.append([example_files["ex3"], "Expanding upwards from the subject", 1024, 1024, "Bottom"])
+     if default_image_path is None: default_image_path = example_files["ex3"]
+
+if not example_list:
+    print("Warning: No example images found in ./examples/. Examples section will be empty.")
+    # Optionally create a placeholder image
+    # placeholder = Image.new('RGB', (512, 512), color = 'grey')
+    # placeholder_path = "./examples/placeholder.png"
+    # placeholder.save(placeholder_path)
+    # example_list.append([placeholder_path, "Placeholder", 1024, 1024, "Middle"])
+    # default_image_path = placeholder_path
+
+# --- UI ---
+with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo: # Added a theme
+    gr.HTML(title)
+
+    with gr.Row():
+        with gr.Column(scale=1): # Left column for inputs
+            input_image = gr.Image(
+                value=default_image_path, # Load default example
+                type="pil",
+                label="Input Image",
+                elem_id="input-image"
+            )
+
+            prompt_input = gr.Textbox(label="Prompt", placeholder="Describe the scene to expand (optional but recommended)...", lines=2)
+
+            with gr.Row():
+                target_ratio = gr.Radio(
+                    label="Target Aspect Ratio",
+                    choices=["9:16", "16:9", "1:1", "Custom"],
+                    value="9:16",
+                    scale=2
+                )
+                alignment_dropdown = gr.Dropdown(
+                    choices=["Middle", "Left", "Right", "Top", "Bottom"],
+                    value="Middle",
+                    label="Align Source Image",
+                    scale=1
                 )
 
+            with gr.Accordion(label="Advanced settings", open=False) as settings_panel:
                 with gr.Row():
-                    with gr.Column(scale=2):
-                        prompt_input = gr.Textbox(label="Prompt (Optional)", placeholder="Describe the scene to expand...")
-                    with gr.Column(scale=1):
-                        run_button = gr.Button("Generate", variant="primary") # Make primary
+                     width_slider = gr.Slider(
+                         label="Target Width", minimum=512, maximum=2048, step=64, value=720
+                     )
+                     height_slider = gr.Slider(
+                         label="Target Height", minimum=512, maximum=2048, step=64, value=1280
+                     )
+                num_inference_steps = gr.Slider(
+                    label="Steps (TCD/Lightning: 1-8)", minimum=1, maximum=12, step=1, value=4
+                )
 
-                with gr.Row():
-                    target_ratio = gr.Radio(
-                        label="Target Ratio",
-                        choices=["9:16", "16:9", "1:1", "Custom"],
-                        value="9:16",
-                        scale=2
+                with gr.Group():
+                    overlap_percentage = gr.Slider(
+                        label="Mask Overlap with Source (%)", minimum=0, maximum=50, value=12, step=1
                     )
+                    gr.Markdown("Select edges to overlap:", scale=0) # Add context
+                    with gr.Row(elem_classes="gradio-checkboxgroup"): # Apply CSS class
+                        overlap_top = gr.Checkbox(label="Top", value=True, scale=1)
+                        overlap_bottom = gr.Checkbox(label="Bottom", value=True, scale=1)
+                        overlap_left = gr.Checkbox(label="Left", value=True, scale=1)
+                        overlap_right = gr.Checkbox(label="Right", value=True, scale=1)
 
-                    alignment_dropdown = gr.Dropdown(
-                        choices=["Middle", "Left", "Right", "Top", "Bottom"],
-                        value="Middle",
-                        label="Align Source Image"
-                    )
 
-                with gr.Accordion(label="Advanced settings", open=False) as settings_panel:
-                    with gr.Row():
-                         width_slider = gr.Slider(
-                             label="Target Width",
-                             minimum=512, # Lowered minimum slightly
-                             maximum=2048, # Increased maximum slightly
-                             step=64, # Use steps of 64 common for SD
-                             value=720,
-                         )
-                         height_slider = gr.Slider(
-                             label="Target Height",
-                             minimum=512,
-                             maximum=2048,
-                             step=64,
-                             value=1280,
-                         )
-                    num_inference_steps = gr.Slider(label="Steps", minimum=1, maximum=12, step=1, value=4) # TCD/Lightning allows few steps
-
-                    with gr.Group():
-                        overlap_percentage = gr.Slider(
-                            label="Mask overlap (%)",
-                            minimum=1,
-                            maximum=50,
-                            value=12, # Default overlap
-                            step=1
-                        )
-                        with gr.Row():
-                            overlap_top = gr.Checkbox(label="Top", value=True)
-                            overlap_right = gr.Checkbox(label="Right", value=True)
-                            overlap_bottom = gr.Checkbox(label="Bottom", value=True)
-                            overlap_left = gr.Checkbox(label="Left", value=True)
-
-
-                    with gr.Row():
-                        resize_option = gr.Radio(
-                            label="Resize input within target",
-                            choices=["Full", "50%", "33%", "25%", "Custom"],
-                            value="Full"
-                        )
-                        custom_resize_percentage = gr.Slider(
-                            label="Custom resize (%)",
-                            minimum=1,
-                            maximum=100,
-                            step=1,
-                            value=50,
-                            visible=False # Initially hidden
-                        )
-
-                    preview_button = gr.Button("Preview Mask & Alignment")
-                    preview_image = gr.Image(label="Mask Preview (Red = Outpaint Area)", type="pil", interactive=False)
+                with gr.Row():
+                    resize_option = gr.Radio(
+                        label="Resize source within target",
+                        choices=["Full", "50%", "33%", "25%", "Custom"],
+                        value="Full",
+                         scale=2
+                    )
+                    custom_resize_percentage = gr.Slider(
+                        label="Custom resize (%)", minimum=1, maximum=100, step=1, value=50, visible=False, scale=1
+                    )
 
+                preview_button = gr.Button("Preview Mask & Alignment")
+                preview_image = gr.Image(label="Mask Preview (Red = Outpaint Area)", type="pil", interactive=False, elem_id="preview-image")
 
+            if example_list:
                 gr.Examples(
-                    examples=[
-                        ["./examples/example_1.webp", "A wide landscape view of the mountains", 1280, 720, "Middle"],
-                        ["./examples/example_2.jpg", "Full body shot of the astronaut on the moon", 720, 1280, "Middle"],
-                        ["./examples/example_3.jpg", "Expanding the sky and ground around the subject", 1024, 1024, "Middle"],
-                         ["./examples/example_3.jpg", "Expanding downwards from the subject", 1024, 1024, "Top"], # Align subject Top
-                         ["./examples/example_3.jpg", "Expanding upwards from the subject", 1024, 1024, "Bottom"], # Align subject Bottom
-                    ],
+                    examples=example_list,
                     inputs=[input_image, prompt_input, width_slider, height_slider, alignment_dropdown],
-                    label="Examples (Click to load)"
+                    label="Examples (Click to load)",
+                    examples_per_page=10
                 )
+            else:
+                 gr.Markdown("_(No example files found in ./examples)_")
 
+            run_button = gr.Button("Generate", variant="primary")
 
-            with gr.Column(scale=1): # Right column for output
-                # Replace ImageSlider with gr.Image
-                result = gr.Image(label="Generated Image", type="pil", interactive=False, elem_id="result-image")
-                use_as_input_button = gr.Button("Use Result as Input Image", visible=False) # Initially hidden
 
-                history_gallery = gr.Gallery(
-                    label="History",
-                    columns=6,
-                    object_fit="contain",
-                    interactive=False,
-                    height=110, # Fixed height for the row
-                    elem_id="history-gallery"
-                )
+        with gr.Column(scale=1): # Right column for output
+            result = gr.Image(label="Generated Image", type="pil", interactive=False, elem_id="result-image")
+            use_as_input_button = gr.Button("Use Result as Input Image", visible=False)
 
+            history_gallery = gr.Gallery(
+                label="History", columns=6, object_fit="contain", interactive=False,
+                height=110, elem_id="history-gallery"
+            )
 
     # --- Event Handling ---
 
-    def use_output_as_input(output_image):
-        """Sets the generated output as the new input image."""
-        # output_image is now the single final image from gr.Image
-        return gr.update(value=output_image)
+    # Function to set result as input and clear result area
+    def use_output_as_input_and_clear(output_image):
+        return {
+            input_image: gr.update(value=output_image),
+            result: gr.update(value=None), # Clear result after using it
+            use_as_input_button: gr.update(visible=False) # Hide button again
+        }
 
     use_as_input_button.click(
-        fn=use_output_as_input,
-        inputs=[result], # Input is the result image component
-        outputs=[input_image] # Output updates the input image component
+        fn=use_output_as_input_and_clear,
+        inputs=[result],
+        outputs=[input_image, result, use_as_input_button]
     )
 
     target_ratio.change(
         fn=preload_presets,
         inputs=[target_ratio, width_slider, height_slider],
-        outputs=[width_slider, height_slider, settings_panel], # Also control accordion state
+        outputs=[width_slider, height_slider, settings_panel],
         queue=False
     )
 
-    # Link sliders back to the ratio selector
     width_slider.change(
         fn=select_the_right_preset,
         inputs=[width_slider, height_slider],
@@ -472,58 +674,71 @@ with gr.Blocks(css=css) as demo:
         resize_option, custom_resize_percentage, prompt_input, alignment_dropdown,
         overlap_left, overlap_right, overlap_top, overlap_bottom
     ]
+    gen_outputs = [result] # Single output image
 
-    # Chain generation logic
-    run_button.click(
-        fn=clear_result,
-        inputs=None,
-        outputs=[result], # Clear the single image output
-        queue=False # Run clearing immediately
+    # Chain generation logic for Run button
+    run_trigger = run_button.click(
+        fn=clear_result, # Clear previous result first
+        inputs=[], # No inputs needed for clear
+        outputs=[result, use_as_input_button], # Components to clear/hide
+        queue=False
     ).then(
         fn=infer,
         inputs=gen_inputs,
-        outputs=[result], # Output the single image to the result component
-    ).then(
-        # Update history with the single result image
+        outputs=gen_outputs,
+    )
+
+    # After generation finishes (successfully or not), update history and button visibility
+    run_trigger.then(
         fn=lambda res_img, hist: update_history(res_img, hist),
         inputs=[result, history_gallery],
         outputs=[history_gallery],
-        queue=False # Update history immediately after generation
+        queue=False # Update history immediately
     ).then(
-        # Show the 'Use as Input' button
-        fn=lambda: gr.update(visible=True),
-        inputs=None,
+        # Show the 'Use as Input' button only if generation was successful (result is not None)
+        fn=lambda res_img: gr.update(visible=isinstance(res_img, Image.Image)),
+        inputs=[result],
         outputs=[use_as_input_button],
         queue=False # Show button immediately
     )
 
-    prompt_input.submit(
+
+    # Chain generation logic for Enter key in Prompt textbox
+    submit_trigger = prompt_input.submit(
          fn=clear_result,
-        inputs=None,
-        outputs=[result],
+        inputs=[],
+        outputs=[result, use_as_input_button],
         queue=False
     ).then(
         fn=infer,
         inputs=gen_inputs,
-        outputs=[result],
-    ).then(
+        outputs=gen_outputs,
+    )
+
+    submit_trigger.then(
         fn=lambda res_img, hist: update_history(res_img, hist),
         inputs=[result, history_gallery],
         outputs=[history_gallery],
          queue=False
     ).then(
-        fn=lambda: gr.update(visible=True),
-        inputs=None,
+        fn=lambda res_img: gr.update(visible=isinstance(res_img, Image.Image)),
+        inputs=[result],
         outputs=[use_as_input_button],
         queue=False
     )
 
+    # Preview button logic
+    preview_inputs = [
+        input_image, width_slider, height_slider, overlap_percentage, resize_option,
+        custom_resize_percentage, alignment_dropdown, overlap_left, overlap_right,
+        overlap_top, overlap_bottom
+    ]
     preview_button.click(
         fn=preview_image_and_mask,
-        inputs=[input_image, width_slider, height_slider, overlap_percentage, resize_option, custom_resize_percentage, alignment_dropdown,
-                overlap_left, overlap_right, overlap_top, overlap_bottom],
+        inputs=preview_inputs,
         outputs=preview_image,
-        queue=False # Preview should be fast
+        queue=False
     )
 
-demo.queue(max_size=10).launch(ssr_mode=False, show_error=True) # Removed share=False for potential Hugging Face Spaces use
+# Launch the interface
+demo.queue(max_size=10).launch(ssr_mode=False, show_error=True, debug=True) # Add debug=True for more logs