Create app.py

app.py

@@ -0,0 +1,143 @@
+import gradio as gr
+import torch
+import imageio
+import os
+import time
+import random
+import gc
+from PIL import Image
+
+# Import necessary components from the cloned repository
+from skyreels_v2_infer.modules import download_model
+from skyreels_v2_infer.pipelines import Image2VideoPipeline, resizecrop
+
+# --- Global Configuration & Model Loading ---
+MODEL_ID = "Skywork/SkyReels-V2-I2V-14B-720P"
+HEIGHT = 720
+WIDTH = 1280
+OUTPUT_DIR = "video_out"
+
+# Create output directory
+os.makedirs(OUTPUT_DIR, exist_ok=True)
+
+print("Downloading and loading model... This may take a while.")
+# Download model files to the cache
+cached_model_path = download_model(MODEL_ID)
+
+# Load the pipeline. This is done once when the Space starts.
+# We enable offload by default to be compatible with GPUs like A10G-Large (24GB VRAM)
+pipe = Image2VideoPipeline(
+    model_path=cached_model_path,
+    dit_path=cached_model_path,
+    use_usp=False,
+    offload=True  # Enable CPU offload to save VRAM
+)
+print("Model loaded successfully.")
+
+# --- Inference Function ---
+def generate_video(input_image, prompt, guidance_scale, inference_steps, num_frames, fps, seed):
+    """
+    Main function to generate video from an image and a prompt.
+    """
+    if input_image is None:
+        raise gr.Error("You must upload an initial image.")
+    if not prompt:
+        raise gr.Error("Prompt cannot be empty.")
+
+    # Use provided seed or generate a random one
+    if seed == -1:
+        seed = random.randint(0, 2**32 - 1)
+    
+    generator = torch.Generator(device="cuda").manual_seed(seed)
+    
+    # Prepare the input image (resize and crop)
+    image = Image.fromarray(input_image).convert("RGB")
+    processed_image = resizecrop(image, HEIGHT, WIDTH)
+
+    # Define a default negative prompt
+    negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, worst quality, low quality, JPEG compression residue, ugly, deformed."
+
+    # Set up generation parameters
+    kwargs = {
+        "image": processed_image,
+        "prompt": prompt,
+        "negative_prompt": negative_prompt,
+        "num_frames": num_frames,
+        "num_inference_steps": inference_steps,
+        "guidance_scale": guidance_scale,
+        "shift": 8.0, # Default value from original script
+        "generator": generator,
+        "height": HEIGHT,
+        "width": WIDTH,
+    }
+    
+    print(f"Generating video with seed: {seed}")
+    start_time = time.time()
+    
+    # Run inference
+    with torch.cuda.amp.autocast(dtype=pipe.transformer.dtype), torch.no_grad():
+        video_frames = pipe(**kwargs)[0]
+
+    end_time = time.time()
+    print(f"Inference took {end_time - start_time:.2f} seconds.")
+
+    # Save the output video
+    # Sanitize prompt for filename
+    safe_prompt = "".join(c for c in prompt if c.isalnum() or c in " _-").strip()[:50]
+    output_filename = f"{safe_prompt}_{seed}.mp4"
+    output_path = os.path.join(OUTPUT_DIR, output_filename)
+    
+    imageio.mimwrite(output_path, video_frames, fps=fps, quality=8, output_params=["-loglevel", "error"])
+    
+    print(f"Video saved to {output_path}")
+    
+    # Clean up memory
+    gc.collect()
+    torch.cuda.empty_cache()
+    
+    return output_path
+
+# --- Gradio UI ---
+with gr.Blocks(css="footer {display: none !important}") as demo:
+    gr.Markdown(
+        """
+        # SkyReels-V2 Image-to-Video Generator
+        ### Model: Skywork/SkyReels-V2-I2V-14B-720P
+        This Space demonstrates the SkyReels V2 model for generating video from a single starting image and a text prompt.
+        **Note:** This is a very large model. Generation can take several minutes, even on powerful GPUs.
+        """
+    )
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(type="numpy", label="Initial Image")
+            prompt = gr.Textbox(label="Prompt", placeholder="e.g., A cinematic shot of a car driving on a rainy street at night.")
+            
+            with gr.Accordion("Advanced Settings", open=False):
+                guidance_scale = gr.Slider(minimum=1.0, maximum=15.0, value=6.0, step=0.5, label="Guidance Scale")
+                inference_steps = gr.Slider(minimum=10, maximum=100, value=30, step=1, label="Inference Steps")
+                num_frames = gr.Slider(minimum=25, maximum=145, value=97, step=8, label="Number of Frames")
+                fps = gr.Slider(minimum=8, maximum=30, value=24, step=1, label="Frames Per Second (FPS)")
+                seed = gr.Number(value=-1, label="Seed (-1 for random)")
+        
+        with gr.Column():
+            output_video = gr.Video(label="Generated Video")
+            run_button = gr.Button("Generate Video", variant="primary")
+
+    gr.Examples(
+        examples=[
+            ["./examples/car.png", "A cinematic shot of a car driving on a rainy street at night, neon lights reflecting on the wet pavement.", 7.0, 30, 97, 24, 12345],
+            ["./examples/castle.png", "An epic fantasy castle in the mountains, dragons flying in the sky, cinematic lighting.", 6.0, 40, 97, 12, 54321],
+        ],
+        inputs=[input_image, prompt, guidance_scale, inference_steps, num_frames, fps, seed],
+        outputs=output_video,
+        fn=generate_video,
+        cache_examples=False, # Set to True if you have GPU and want to pre-process examples
+    )
+
+    # Add example images to your space in a folder named 'examples' for this to work
+    # Or simply remove the gr.Examples block.
+
+    run_button.click(fn=generate_video, inputs=[input_image, prompt, guidance_scale, inference_steps, num_frames, fps, seed], outputs=output_video)
+
+if __name__ == "__main__":
+    demo.launch()