Generate creative video prompts with technical specifications
You can use prompts with Kling, MiniMax, Hunyuan, Haiper, CogVideoX, Luma, LTX, Runway, PixVerse.
""" # Import these at global scope but don't instantiate yet from vlm_captions import VLMCaptioning from llm_inference_video import VideoLLMInferenceNode # Global singleton instances - we'll initialize them only when needed vlm_captioner = None llm_node = None # Initialize only once on first use def get_vlm_captioner(): global vlm_captioner if vlm_captioner is None: print("Initializing Video Prompt Generator...") vlm_captioner = VLMCaptioning() print("Video Prompt Generator initialized successfully!") return vlm_captioner def get_llm_node(): global llm_node if llm_node is None: llm_node = VideoLLMInferenceNode() return llm_node # Wrapper functions that avoid passing the model between processes @spaces.GPU() def describe_image_wrapper(image, question="Describe this image in detail."): """GPU-decorated function for image description""" if image is None: return "Please upload an image." if not question or question.strip() == "": question = "Describe this image in detail." # Get the captioner inside this GPU-decorated function vlm = get_vlm_captioner() return vlm.describe_image(image=image, question=question) @spaces.GPU() def describe_video_wrapper(video, frame_interval=30): """GPU-decorated function for video description""" if video is None: return "Please upload a video." # Get the captioner inside this GPU-decorated function vlm = get_vlm_captioner() return vlm.describe_video(video_path=video, frame_interval=frame_interval) def generate_video_prompt_wrapper( concept, style, camera_style, camera_direction, pacing, special_effects, custom_elements, provider, model, prompt_length ): """Wrapper for LLM prompt generation""" node = get_llm_node() return node.generate_video_prompt( concept, style, camera_style, camera_direction, pacing, special_effects, custom_elements, provider, model, prompt_length ) def create_video_interface(): with gr.Blocks(theme='bethecloud/storj_theme') as demo: gr.HTML(title) with gr.Tab("Video Prompt Generator"): with gr.Row(): with gr.Column(scale=1): input_concept = gr.Textbox(label="Core Concept/Thematic Input", lines=3) style = gr.Dropdown( choices=["Minimalist", "Simple", "Detailed", "Descriptive", "Dynamic", "Cinematic", "Documentary", "Animation", "Action", "Experimental"], value="Simple", label="Video Style" ) custom_elements = gr.Textbox(label="Custom Technical Elements", placeholder="e.g., Infrared hybrid, Datamosh transitions") prompt_length = gr.Dropdown( choices=["Short", "Medium", "Long"], value="Medium", label="Prompt Length" ) with gr.Column(scale=1): camera_direction = gr.Dropdown( choices=[ "None", "Zoom in", "Zoom out", "Pan left", "Pan right", "Tilt up", "Tilt down", "Orbital rotation", "Push in", "Pull out", "Track forward", "Track backward", "Spiral in", "Spiral out", "Arc movement", "Diagonal traverse", "Vertical rise", "Vertical descent" ], value="None", label="Camera Direction" ) camera_style = gr.Dropdown( choices=[ "None", "Steadicam flow", "Drone aerials", "Handheld urgency", "Crane elegance", "Dolly precision", "VR 360", "Multi-angle rig", "Static tripod", "Gimbal smoothness", "Slider motion", "Jib sweep", "POV immersion", "Time-slice array", "Macro extreme", "Tilt-shift miniature", "Snorricam character", "Whip pan dynamics", "Dutch angle tension", "Underwater housing", "Periscope lens" ], value="None", label="Camera Movement Style" ) pacing = gr.Dropdown( choices=[ "None", "Slow burn", "Rhythmic pulse", "Frantic energy", "Ebb and flow", "Hypnotic drift", "Time-lapse rush", "Stop-motion staccato", "Gradual build", "Quick cut rhythm", "Long take meditation", "Jump cut energy", "Match cut flow", "Cross-dissolve dreamscape", "Parallel action", "Slow motion impact", "Ramping dynamics", "Montage tempo", "Continuous flow", "Episodic breaks" ], value="None", label="Pacing Rhythm" ) special_effects = gr.Dropdown( choices=[ "None", "Practical effects", "CGI enhancement", "Analog glitches", "Light painting", "Projection mapping", "Nanosecond exposures", "Double exposure", "Smoke diffusion", "Lens flare artistry", "Particle systems", "Holographic overlay", "Chromatic aberration", "Digital distortion", "Wire removal", "Motion capture", "Miniature integration", "Weather simulation", "Color grading", "Mixed media composite", "Neural style transfer" ], value="None", label="SFX Approach" ) with gr.Column(scale=1): provider = gr.Dropdown( choices=["SambaNova", "Groq"], value="SambaNova", label="LLM Provider" ) model = gr.Dropdown( choices=[ "Meta-Llama-3.1-70B-Instruct", "Meta-Llama-3.1-405B-Instruct", "Meta-Llama-3.1-8B-Instruct" ], value="Meta-Llama-3.1-70B-Instruct", label="Model" ) generate_btn = gr.Button("Generate Video Prompt", variant="primary") output = gr.Textbox(label="Generated Prompt", lines=12, show_copy_button=True) def update_models(provider): models = { "Groq": ["llama-3.3-70b-versatile"], "SambaNova": [ "Meta-Llama-3.1-70B-Instruct", "Meta-Llama-3.1-405B-Instruct", "Meta-Llama-3.1-8B-Instruct" ] } return gr.Dropdown(choices=models[provider], value=models[provider][0]) provider.change(update_models, inputs=provider, outputs=model) generate_btn.click( generate_video_prompt_wrapper, inputs=[input_concept, style, camera_style, camera_direction, pacing, special_effects, custom_elements, provider, model, prompt_length], outputs=output ) with gr.Tab("Visual Analysis"): with gr.Row(): with gr.Column(): image_input = gr.Image(label="Upload Image", type="filepath") image_question = gr.Textbox( label="Question (optional)", placeholder="What is in this image?" ) analyze_image_btn = gr.Button("Analyze Image") image_output = gr.Textbox(label="Analysis Result", lines=5) with gr.Column(): video_input = gr.Video(label="Upload Video") analyze_video_btn = gr.Button("Analyze Video") video_output = gr.Textbox(label="Video Analysis", lines=10) # Use GPU-decorated wrapper functions directly analyze_image_btn.click( describe_image_wrapper, inputs=[image_input, image_question], outputs=image_output ) analyze_video_btn.click( describe_video_wrapper, inputs=video_input, outputs=video_output ) return demo if __name__ == "__main__": demo = create_video_interface() # Don't use share=True on Hugging Face Spaces demo.launch()