import os os.environ["GRADIO_TEMP_DIR"] = os.path.join(os.getcwd(), ".tmp_outputs") os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" import uuid import gradio as gr import spaces from videosys import CogVideoXConfig, CogVideoXPABConfig, VideoSysEngine PROMPT = "A modern living room with a minimalist design, featuring a large window, a white ceiling, and a wooden floor. The room is furnished with a white sofa, a gray ottoman, a wooden table, and a hanging light. The space is well-lit and has a clean, contemporary aesthetic." def load_model(model_name, enable_video_sys=False, pab_threshold=[100, 850], pab_range=2): pab_config = CogVideoXPABConfig(spatial_threshold=pab_threshold, spatial_range=pab_range) config = CogVideoXConfig(model_name, enable_pab=enable_video_sys, pab_config=pab_config) engine = VideoSysEngine(config) return engine def generate(engine, prompt, num_inference_steps=50, guidance_scale=6.0): video = engine.generate(prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale).video[0] unique_filename = f"{uuid.uuid4().hex}.mp4" output_path = os.path.join("./.tmp_outputs", unique_filename) engine.save_video(video, output_path) return output_path @spaces.GPU(duration=300) def generate_vs( model_name, prompt, num_inference_steps, guidance_scale, threshold_start, threshold_end, gap, progress=gr.Progress(track_tqdm=True), ): threshold = [int(threshold_end), int(threshold_start)] gap = int(gap) engine = load_model(model_name, enable_video_sys=True, pab_threshold=threshold, pab_range=gap) video_path = generate(engine, prompt, num_inference_steps, guidance_scale) return video_path css = """ body { font-family: Arial, sans-serif; line-height: 1.6; color: #333; margin: 0 auto; padding: 20px; } .container { display: flex; flex-direction: column; gap: 10px; } .row { display: flex; flex-wrap: wrap; gap: 10px; } .column { flex: 1; min-width: 0; } .video-output { width: 100%; max-width: 720px; height: auto; margin: 0 auto; } .server-status { margin-top: 5px; padding: 5px; font-size: 0.8em; } .server-status h4 { margin: 0 0 3px 0; font-size: 0.9em; } .server-status .row { margin-bottom: 2px; } .server-status .textbox { min-height: unset !important; } .server-status .textbox input { padding: 1px 5px !important; height: 20px !important; font-size: 0.9em !important; } .server-status .textbox label { margin-bottom: 0 !important; font-size: 0.9em !important; line-height: 1.2 !important; } .server-status .textbox { gap: 0 !important; } .server-status .textbox input { margin-top: -2px !important; } @media (max-width: 768px) { .row { flex-direction: column; } .column { width: 100%; } } .video-output { width: 100%; height: auto; } } """ with gr.Blocks(css=css) as demo: gr.HTML( """
KoolCogVideoX Huggingface Space🤗
KoolCogVideoX is fine-tuned on CogVideoX specifically for interior design scenarios.
The demo is powered by https://github.com/NUS-HPC-AI-Lab/VideoSys.
⚠️ This demo is for academic research and experiential use only. Users should strictly adhere to local laws and ethics.
""" ) with gr.Row(): with gr.Column(): prompt = gr.Textbox(label="Prompt (Less than 200 Words)", value=PROMPT, lines=2) with gr.Column(): gr.Markdown("**Generation Parameters**
") with gr.Row(): model_name = gr.Radio( ["bertjiazheng/KoolCogVideoX-2b", "bertjiazheng/KoolCogVideoX-5b"], label="Model Type", value="bertjiazheng/KoolCogVideoX-2b" ) with gr.Row(): num_inference_steps = gr.Slider(label="Inference Steps", maximum=50, value=50) guidance_scale = gr.Slider(label="Guidance Scale", value=6.0, maximum=15.0) gr.Markdown("**Pyramid Attention Broadcast Parameters**
") with gr.Row(): pab_range = gr.Slider( label="Broadcast Range", value=2, step=1, minimum=1, maximum=4, info="Attention broadcast range.", ) pab_threshold_start = gr.Slider( label="Start Timestep", minimum=500, maximum=1000, value=850, step=1, info="Broadcast start timestep (1000 is the fisrt).", ) pab_threshold_end = gr.Slider( label="End Timestep", minimum=0, maximum=500, step=1, value=100, info="Broadcast end timestep (0 is the last).", ) with gr.Row(): generate_button_vs = gr.Button("⚡️ Generate Video with VideoSys") with gr.Column(): with gr.Row(): video_output_vs = gr.Video(label="CogVideoX with VideoSys", width=720, height=480) gr.Markdown("""
🎥 Video Gallery
These videos are generated by KoolCogVideoX-5b.

A modern living room with a minimalist design, featuring a white sofa, a marble coffee table, a geometric painting, and a chandelier hanging from the ceiling. The room is well-lit with natural light, and the color scheme is neutral with accents of gold and black. The furniture is arranged in a way that creates a comfortable and inviting space.

A modern living room with a minimalist design, featuring a large window, a white ceiling, and a wooden floor. The room is furnished with a white sofa, a gray ottoman, a wooden table, and a hanging light. The space is well-lit and has a clean, contemporary aesthetic.

A modern bedroom with a minimalist design, featuring a large bed with a gray comforter and a blue blanket, a white dresser with a mirror, and a white closet. The room is decorated with framed artwork and a black and white poster on the wall. The floor is made of light wood, and the room has a clean and contemporary feel.

A modern kitchen with a sleek design, featuring a marble countertop, stainless steel appliances, and a variety of bottles and glasses. The kitchen is well-lit with recessed lighting and has a contemporary aesthetic.

""") generate_button_vs.click( generate_vs, inputs=[ model_name, prompt, num_inference_steps, guidance_scale, pab_threshold_start, pab_threshold_end, pab_range, ], outputs=[video_output_vs], concurrency_id="gen", concurrency_limit=1, ) if __name__ == "__main__": demo.queue(max_size=10, default_concurrency_limit=1) demo.launch()