By Gemini

app.py

@@ -4,54 +4,173 @@ import random
 import spaces
 import torch
 import time
+import logging
 from diffusers import DiffusionPipeline, AutoencoderTiny
+# Using AttnProcessor2_0 for potential speedup with PyTorch 2.x
 from diffusers.models.attention_processor import AttnProcessor2_0
+# Assuming custom_pipeline defines FluxWithCFGPipeline correctly
 from custom_pipeline import FluxWithCFGPipeline
 
+# --- Setup Logging ---
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+# --- Torch Optimizations ---
 torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.benchmark = True # Enable cuDNN benchmark for potentially faster convolutions
 
-# Constants
+# --- Constants ---
 MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 2048
+MAX_IMAGE_SIZE = 2048 # Keep a reasonable limit to prevent OOMs
 DEFAULT_WIDTH = 1024
 DEFAULT_HEIGHT = 1024
-DEFAULT_INFERENCE_STEPS = 1
+DEFAULT_INFERENCE_STEPS = 1 # FLUX Schnell is designed for few steps
+MIN_INFERENCE_STEPS = 1
+MAX_INFERENCE_STEPS = 8 # Allow slightly more steps for potential quality boost
+ENHANCE_STEPS = 4 # Fixed steps for the enhance button
 
-# Device and model setup
+# --- Device and Model Setup ---
 dtype = torch.float16
-pipe = FluxWithCFGPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-schnell", torch_dtype=dtype
-)
-pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype)
-pipe.to("cuda")
-pipe.load_lora_weights('hugovntr/flux-schnell-realism', weight_name='schnell-realism_v2.3.safetensors', adapter_name="better")
-pipe.set_adapters(["better"], adapter_weights=[1.0])
-pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0)
-pipe.unload_lora_weights()
-
-torch.cuda.empty_cache()
-
-# Inference function
-@spaces.GPU(duration=25)
-def generate_image(prompt, seed=24, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT, randomize_seed=False, num_inference_steps=2, progress=gr.Progress(track_tqdm=True)):
+device = "cuda" if torch.cuda.is_available() else "cpu"
+pipe = None # Initialize pipe to None
+
+try:
+    logging.info("Loading diffusion pipeline...")
+    pipe = FluxWithCFGPipeline.from_pretrained(
+        "black-forest-labs/FLUX.1-schnell", torch_dtype=dtype
+    )
+    logging.info("Loading VAE...")
+    pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype)
+    
+    logging.info(f"Moving pipeline to {device}...")
+    pipe.to(device)
+
+    # Apply optimizations
+    logging.info("Setting attention processor...")
+    pipe.unet.set_attn_processor(AttnProcessor2_0())
+    pipe.vae.set_attn_processor(AttnProcessor2_0()) # VAE might benefit too
+
+    logging.info("Loading and fusing LoRA...")
+    pipe.load_lora_weights('hugovntr/flux-schnell-realism', weight_name='schnell-realism_v2.3.safetensors', adapter_name="better")
+    pipe.set_adapters(["better"], adapter_weights=[1.0])
+    pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0) # Fuse for potential speedup
+    pipe.unload_lora_weights() # Unload after fusing
+    logging.info("LoRA fused and unloaded.")
+
+    # --- Compilation (Major Speed Optimization) ---
+    # Note: Compilation takes time on the first run.
+    # logging.info("Compiling UNet (this may take a moment)...")
+    # pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True) # Use reduce-overhead for dynamic shapes
+    # logging.info("Compiling VAE Decoder...")
+    # pipe.vae.decoder = torch.compile(pipe.vae.decoder, mode="reduce-overhead", fullgraph=True)
+    # logging.info("Compiling VAE Encoder...")
+    # pipe.vae.encoder = torch.compile(pipe.vae.encoder, mode="reduce-overhead", fullgraph=True)
+    # logging.info("Model compilation finished.")
+
+    # --- Optional: Warm-up Run ---
+    # logging.info("Performing warm-up run...")
+    # with torch.inference_mode():
+    #     _ = pipe(prompt="warmup", num_inference_steps=1, generator=torch.Generator(device=device).manual_seed(0), output_type="pil", return_dict=False)[0]
+    # logging.info("Warm-up complete.")
+
+    # Clear cache after setup
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        logging.info("CUDA cache cleared after setup.")
+
+except Exception as e:
+    logging.error(f"Error during model loading or setup: {e}", exc_info=True)
+    # Display error in Gradio if UI is already built, otherwise just log and exit.
+    # For simplicity here, we'll rely on the Gradio UI showing an error if `pipe` is None later.
+    # If running script directly, consider `sys.exit()`
+    # raise gr.Error(f"Failed to load models. Check logs for details. Error: {e}")
+
+
+# --- Inference Function ---
+@spaces.GPU(duration=30) # Slightly increased duration buffer
+def generate_image(prompt: str, seed: int = 42, width: int = DEFAULT_WIDTH, height: int = DEFAULT_HEIGHT, randomize_seed: bool = False, num_inference_steps: int = DEFAULT_INFERENCE_STEPS, is_enhance: bool = False):
+    """Generates an image using the FLUX pipeline with error handling."""
+    
+    if pipe is None:
+        raise gr.Error("Diffusion pipeline failed to load. Cannot generate images.")
+        
+    if not prompt or prompt.strip() == "":
+         # Return a blank image or previous result if prompt is empty?
+         # For now, raise warning and return None.
+         gr.Warning("Prompt is empty. Please enter a description.")
+         # Returning None for image, original seed, and error message
+         return None, seed, "Error: Empty prompt"
+
+    start_time = time.time()
+    
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator().manual_seed(int(float(seed)))
+    
+    # Clamp dimensions to avoid excessive memory usage
+    width = min(width, MAX_IMAGE_SIZE)
+    height = min(height, MAX_IMAGE_SIZE)
+    
+    # Use fixed steps for enhance button, otherwise use slider value
+    steps_to_use = ENHANCE_STEPS if is_enhance else num_inference_steps
+    # Clamp steps
+    steps_to_use = max(MIN_INFERENCE_STEPS, min(steps_to_use, MAX_INFERENCE_STEPS))
 
-    start_time = time.time()
+    logging.info(f"Generating image with prompt: '{prompt}', seed: {seed}, size: {width}x{height}, steps: {steps_to_use}")
+
+    try:
+        # Ensure generator is on the correct device
+        generator = torch.Generator(device=device).manual_seed(int(float(seed)))
+
+        # Use inference_mode for efficiency
+        with torch.inference_mode():
+            # Generate the image (assuming pipe returns list/tuple with image first)
+            # Modify pipe call based on its actual signature if needed
+            result_img = pipe(
+                prompt=prompt,
+                width=width,
+                height=height,
+                num_inference_steps=steps_to_use,
+                generator=generator,
+                output_type="pil", # Ensure PIL output for Gradio Image component
+                return_dict=False # Assuming the custom pipeline supports this for direct output
+            )[0][0] # Assuming the output structure is [[img]]
+
+        latency = time.time() - start_time
+        latency_str = f"Latency: {latency:.2f} seconds (Steps: {steps_to_use})"
+        logging.info(f"Image generated successfully. {latency_str}")
+        return result_img, seed, latency_str
+
+    except torch.cuda.OutOfMemoryError as e:
+        logging.error(f"CUDA OutOfMemoryError: {e}", exc_info=True)
+        # Clear cache and suggest reducing size/steps
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        raise gr.Error("GPU ran out of memory. Try reducing the image width/height or the number of inference steps.")
+        
+    except Exception as e:
+        logging.error(f"Error during image generation: {e}", exc_info=True)
+        # Clear cache just in case
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        raise gr.Error(f"An error occurred during generation: {e}")
 
-    # Only generate the last image in the sequence
-    img = pipe.generate_images( 
-            prompt=prompt,
-            width=width,
-            height=height,
-            num_inference_steps=num_inference_steps,
-            generator=generator
-        )
-    latency = f"Latency: {(time.time()-start_time):.2f} seconds"    
-    return img, seed, latency
 
-# Example prompts
+# --- Real-time Generation Wrapper ---
+# This function checks the realtime toggle before calling the main generation function.
+# It's triggered by changes in prompt or sliders when realtime is enabled.
+def handle_realtime_update(realtime_enabled: bool, prompt: str, seed: int, width: int, height: int, randomize_seed: bool, num_inference_steps: int):
+    if realtime_enabled and pipe is not None:
+        logging.debug("Realtime update triggered.")
+        # Call generate_image directly. Errors within generate_image will be caught and raised as gr.Error.
+        # We don't set is_enhance=True for realtime updates.
+        return generate_image(prompt, seed, width, height, randomize_seed, num_inference_steps, is_enhance=False)
+    else:
+        # If realtime is disabled or pipe failed, don't update the image, seed, or latency.
+        # Return gr.update() for each output component to indicate no change.
+        logging.debug("Realtime update skipped (disabled or pipe error).")
+        return gr.update(), gr.update(), gr.update()
+
+
+# --- Example Prompts ---
 examples = [
     "a tiny astronaut hatching from an egg on the moon",
     "a cute white cat holding a sign that says hello world",
@@ -60,107 +179,130 @@ examples = [
     "photo of a woman on the beach, shot from above. She is facing the sea, while wearing a white dress. She has long blonde hair",
     "Selfie photo of a wizard with long beard and purple robes, he is apparently in the middle of Tokyo. Probably taken from a phone.",
     "Photo of a young woman with long, wavy brown hair tied in a bun and glasses. She has a fair complexion and is wearing subtle makeup, emphasizing her eyes and lips. She is dressed in a black top. The background appears to be an urban setting with a building facade, and the sunlight casts a warm glow on her face.",
+    "High-resolution photorealistic render of a sleek, futuristic motorcycle parked on a neon-lit street at night, rain reflecting the lights.",
+    "Watercolor painting of a cozy bookstore interior with overflowing shelves and a cat sleeping in a sunbeam.",
 ]
 
 # --- Gradio UI ---
-with gr.Blocks() as demo:
-    with gr.Column(elem_id="app-container"):
-        gr.Markdown("# 🎨 Realtime FLUX Image Generator")
-        gr.Markdown("Generate stunning images in real-time with Modified Flux.Schnell pipeline.")
-        gr.Markdown("<span style='color: red;'>Note: Sometimes it stucks or stops generating images (I don't know why). In that situation just refresh the site.</span>")
-
-        with gr.Row():
-            with gr.Column(scale=2.5):
-                result = gr.Image(label="Generated Image", show_label=False, interactive=False)
-            with gr.Column(scale=1):
-                prompt = gr.Text(
-                    label="Prompt",
-                    placeholder="Describe the image you want to generate...",
-                    lines=3,
-                    show_label=False,
-                    container=False,
-                )
-                generateBtn = gr.Button("🖼️ Generate Image")
-                enhanceBtn = gr.Button("🚀 Enhance Image")
-
-                with gr.Column("Advanced Options"):
-                    with gr.Row():
-                        realtime = gr.Checkbox(label="Realtime Toggler", info="If TRUE then uses more GPU but create image in realtime.", value=False)
-                        latency = gr.Text(label="Latency")
-                    with gr.Row():
-                        seed = gr.Number(label="Seed", value=42)
-                        randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
-                    with gr.Row():
-                        width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=DEFAULT_WIDTH)
-                        height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=DEFAULT_HEIGHT)
-                        num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=4, step=1, value=DEFAULT_INFERENCE_STEPS)
-
-        with gr.Row():
-            gr.Markdown("### 🌟 Inspiration Gallery")
-        with gr.Row():
-            gr.Examples(
-                examples=examples,
-                fn=generate_image,
-                inputs=[prompt],
-                outputs=[result, seed, latency],
-                cache_examples="lazy" 
+with gr.Blocks(css="#app-container { max-width: 1280px; margin: auto; }") as demo:
+    gr.Markdown("# 🎨 Realtime FLUX Image Generator")
+    gr.Markdown("Generate stunning images in real-time with Modified Flux.Schnell pipeline. Optimized for speed.")
+    gr.Markdown("<span style='color: red;'>Note: Realtime generation requires a capable GPU. If generation stops or fails, try refreshing or reducing image size/steps.</span>")
+
+    if pipe is None:
+        gr.Markdown("<h2 style='color: red; text-align: center;'>Critical Error: Failed to load models. The application cannot function. Please check the logs.</h2>")
+
+    with gr.Row():
+        with gr.Column(scale=3): # Give image slightly more space
+            result = gr.Image(label="Generated Image", show_label=False, interactive=False, height=768) # Adjust height as needed
+            latency = gr.Text(label="Generation Info", interactive=False)
+        
+        with gr.Column(scale=2):
+            prompt = gr.Text(
+                label="Prompt",
+                placeholder="Describe the image you want to generate...",
+                lines=3,
+                show_label=False,
+                container=False,
             )
+            
+            with gr.Row():
+                 generateBtn = gr.Button("🖼️ Generate Image", variant="primary", interactive=pipe is not None)
+                 enhanceBtn = gr.Button(f"🚀 Enhance (Steps: {ENHANCE_STEPS})", interactive=pipe is not None) # Use fixed steps for enhance
+            
+            realtime = gr.Checkbox(label="⚡ Realtime Generation", info="Generates image automatically as you type or adjust sliders (requires more GPU).", value=False, interactive=pipe is not None)
 
-    enhanceBtn.click(
-        fn=generate_image,
-        inputs=[prompt, seed, width, height],
-        outputs=[result, seed, latency],
-        show_progress="full",
-        queue=False,
-        concurrency_limit=None
+            with gr.Accordion("Advanced Options", open=False):
+                with gr.Row():
+                    seed = gr.Number(label="Seed", value=42, precision=0, interactive=pipe is not None) # Use precision=0 for integers
+                    randomize_seed = gr.Checkbox(label="Randomize Seed", value=True, interactive=pipe is not None)
+                with gr.Row():
+                    width = gr.Slider(label="Width", minimum=256, maximum=MAX_IMAGE_SIZE, step=64, value=DEFAULT_WIDTH, interactive=pipe is not None) # Increase step for faster adjustment
+                    height = gr.Slider(label="Height", minimum=256, maximum=MAX_IMAGE_SIZE, step=64, value=DEFAULT_HEIGHT, interactive=pipe is not None)
+                num_inference_steps = gr.Slider(
+                    label="Inference Steps", 
+                    minimum=MIN_INFERENCE_STEPS, 
+                    maximum=MAX_INFERENCE_STEPS, 
+                    step=1, 
+                    value=DEFAULT_INFERENCE_STEPS, 
+                    info=f"Controls quality vs speed. Default: {DEFAULT_INFERENCE_STEPS}. Enhance uses {ENHANCE_STEPS}.",
+                    interactive=pipe is not None
+                )
+
+    gr.Markdown("---") # Separator
+    gr.Markdown("### 🌟 Inspiration Gallery")
+    gr.Examples(
+        examples=examples,
+        fn=generate_image, # Examples directly call generate_image
+        inputs=[prompt], # Only prompt needed for examples, others use defaults
+        outputs=[result, seed, latency], # Match output components
+        cache_examples="lazy", # Use caching
+        run_on_click=True, # Ensure examples run when clicked
+        label="Example Prompts"
     )
 
+    # --- Event Listeners ---
+
+    # Combine inputs needed for generate_image
+    gen_inputs = [prompt, seed, width, height, randomize_seed, num_inference_steps]
+    outputs = [result, seed, latency]
+
+    # Generate Button Click
     generateBtn.click(
         fn=generate_image,
-        inputs=[prompt, seed, width, height, randomize_seed, num_inference_steps],
-        outputs=[result, seed, latency],
-        show_progress="full",
-        api_name="RealtimeFlux",
-        queue=False
+        inputs=gen_inputs,
+        outputs=outputs,
+        show_progress="full", # Show progress for explicit clicks
+        api_name="generate_flux_image",
+        queue=True # Use queue for button clicks to handle multiple requests gracefully
     )
 
-    def update_ui(realtime_enabled):
-        return {
-            prompt: gr.update(interactive=True),
-            generateBtn: gr.update(visible=not realtime_enabled)
-        }
-
-    realtime.change(
-        fn=update_ui,
-        inputs=[realtime],
-        outputs=[prompt, generateBtn],
-        queue=False,
-        concurrency_limit=None
+    # Enhance Button Click - uses fixed steps
+    enhanceBtn.click(
+        fn=generate_image,
+        # Pass is_enhance=True as the last argument
+        inputs=[prompt, seed, width, height, randomize_seed, num_inference_steps, gr.Checkbox(value=True, visible=False)], # Pass True for is_enhance
+        outputs=outputs,
+        show_progress="full",
+        queue=True 
     )
-
-    def realtime_generation(*args):
-        if args[0]:  # If realtime is enabled
-            return next(generate_image(*args[1:]))
-
+    
+    # Prompt Submission (Enter key)
     prompt.submit(
         fn=generate_image,
-        inputs=[prompt, seed, width, height, randomize_seed, num_inference_steps],
-        outputs=[result, seed, latency],
+        inputs=gen_inputs,
+        outputs=outputs,
         show_progress="full",
-        queue=False,
-        concurrency_limit=None
+        queue=True 
     )
 
-    for component in [prompt, width, height, num_inference_steps]:
-        component.input(
-            fn=realtime_generation,
-            inputs=[realtime, prompt, seed, width, height, randomize_seed, num_inference_steps],
-            outputs=[result, seed, latency],
-            show_progress="hidden",
-            trigger_mode="always_last",
-            queue=False,
-            concurrency_limit=None
+    # --- Realtime Updates ---
+    # List of components that trigger realtime updates
+    realtime_triggers = [prompt, width, height, num_inference_steps, seed, randomize_seed]
+    
+    # Inputs for the realtime handler function
+    realtime_inputs = [realtime, prompt, seed, width, height, randomize_seed, num_inference_steps]
+
+    for component in realtime_triggers:
+        # Use 'input' for sliders/text, 'change' for checkboxes/radio if needed
+        # Using .input for text allows updates while typing
+        # Using .change for sliders updates when released (default) or continuously if specified
+        event_type = "input" if isinstance(component, (gr.Textbox, gr.Number)) else "change"
+        
+        getattr(component, event_type)(
+            fn=handle_realtime_update,
+            inputs=realtime_inputs,
+            outputs=outputs,
+            show_progress="hidden", # Hide progress for realtime updates
+            # queue=False essential for responsiveness & cancelling previous requests
+            # trigger_mode='throttle' with a small delay (e.g., 0.5s) can prevent excessive calls
+            # 'always_last' ensures only the latest input value triggers execution
+            queue=False, 
+            trigger_mode="throttle", 
+            throttle_delay=0.5 # Throttle updates slightly (e.g., every 500ms)
+            # trigger_mode="always_last", # Alternative: trigger only after user stops changing for a bit
         )
 
-# Launch the app
-demo.launch()
+# --- Launch the App ---
+if __name__ == "__main__":
+    demo.queue().launch(debug=True) # Enable queue for better handling of multiple users/requests, add debug=True for more logs