import os import gradio as gr from utils.gradio_utils import * import argparse GRADIO_CACHE = "" parser = argparse.ArgumentParser() parser.add_argument('--public_access', action='store_true') args = parser.parse_args() streaming_svd = StreamingSVD(load_argv=False) on_huggingspace = os.environ.get("SPACE_AUTHOR_NAME") == "PAIR" examples = [ ["Experience the dance of jellyfish: float through mesmerizing swarms of jellyfish, pulsating with otherworldly grace and beauty.", "200 - frames (recommended)", 33, None, None], ["Dive into the depths of the ocean: explore vibrant coral reefs, mysterious underwater caves, and the mesmerizing creatures that call the sea home.", "200 - frames (recommended)", 33, None, None], ["A cute cat.", "200 - frames (recommended)", 33, None, None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test1.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test2.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test3.png", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test4.png", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test5.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test6.png", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test7.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test8.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test9.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test10.jpg", None], ["", "200 - frames (recommended)", 33, "__assets__/gradio_cached_examples/test11.jpg", None], ] def generate(prompt, num_frames, seed, image: np.ndarray): if num_frames == [] or num_frames is None: num_frames = 50 else: num_frames = int(num_frames.split(" ")[0]) if num_frames > 200: # and on_huggingspace: num_frames = 200 if image is None: image = text_to_image_gradio( prompt=prompt, streaming_svd=streaming_svd, seed=seed) video_file_stage_one = image_to_video_vfi_gradio( img=image, num_frames=num_frames, streaming_svd=streaming_svd, seed=seed, gradio_cache=GRADIO_CACHE) expanded_size, orig_size, scaled_outpainted_image = retrieve_intermediate_data(video_file_stage_one) video_file_stage_two = enhance_video_vfi_gradio( img=scaled_outpainted_image, video=video_file_stage_one.replace("__cropped__", "__expanded__"), num_frames=24, streaming_svd=streaming_svd, seed=seed, expanded_size=expanded_size, orig_size=orig_size, gradio_cache=GRADIO_CACHE) return image, video_file_stage_one, video_file_stage_two def enhance(prompt, num_frames, seed, image: np.ndarray, video:str): if num_frames == [] or num_frames is None: num_frames = 50 else: num_frames = int(num_frames.split(" ")[0]) if num_frames > 200: # and on_huggingspace: num_frames = 200 # User directly applied Long Video Generation (without preview) with Flux. if image is None: image = text_to_image_gradio( prompt=prompt, streaming_svd=streaming_svd, seed=seed) # User directly applied Long Video Generation (without preview) with or without Flux. if video is None: video = image_to_video_gradio( img=image, num_frames=(num_frames+1) // 2, streaming_svd=streaming_svd, seed=seed, gradio_cache=GRADIO_CACHE) expanded_size, orig_size, scaled_outpainted_image = retrieve_intermediate_data(video) # Here the video is path and image is numpy array video_file_stage_two = enhance_video_vfi_gradio( img=scaled_outpainted_image, video=video.replace("__cropped__", "__expanded__"), num_frames=num_frames, streaming_svd=streaming_svd, seed=seed, expanded_size=expanded_size, orig_size=orig_size, gradio_cache=GRADIO_CACHE) return image, video_file_stage_two with gr.Blocks() as demo: GRADIO_CACHE = demo.GRADIO_CACHE gr.HTML("""

StreamingSVD

A StreamingT2V method for high-quality long video generation

Roberto Henschel1*, Levon Khachatryan1*, Daniil Hayrapetyan1*, Hayk Poghosyan1, Vahram Tadevosyan1, Zhangyang Wang1,2, Shant Navasardyan1, Humphrey Shi1,3

1Picsart AI Resarch (PAIR), 2UT Austin, 3SHI Labs @ Georgia Tech, Oregon & UIUC

*Equal Contribution

[arXiv] [GitHub]

StreamingSVD is an advanced autoregressive technique for text-to-video and image-to-video generation, generating long hiqh-quality videos with rich motion dynamics, turning SVD into a long video generator. Our method ensures temporal consistency throughout the video, aligns closely to the input text/image, and maintains high frame-level image quality. Our demonstrations include successful examples of videos up to 200 frames, spanning 8 seconds, and can be extended for even longer durations.

""") if on_huggingspace: gr.HTML("""

For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
Duplicate Space

""") with gr.Row(): with gr.Column(scale=1): with gr.Row(): with gr.Column(): with gr.Row(): num_frames = gr.Dropdown(["50 - frames (recommended)", "80 - frames (recommended)", "140 - frames (recommended)", "200 - frames (recommended)", "500 - frames", "1000 - frames", "10000 - frames"], label="Number of Video Frames", info="For >200 frames use local workstation!", value="50 - frames (recommended)") with gr.Row(): prompt_stage1 = gr.Textbox(label='Text-to-Video (Enter text prompt here)', interactive=True, max_lines=1) with gr.Row(): image_stage1 = gr.Image(label='Image-to-Video (Upload Image here, text prompt will be ignored for I2V if entered)', show_label=True, show_download_button=True, interactive=True, height=250) with gr.Column(): video_stage1 = gr.Video(label='Long Video Preview', show_label=True, interactive=False, show_download_button=True, height=203) with gr.Row(): run_button_stage1 = gr.Button("Long Video Generation (faster preview)") with gr.Row(): with gr.Column(): with gr.Accordion('Advanced options', open=False): seed = gr.Slider(label='Seed', minimum=0, maximum=65536, value=33, step=1,) with gr.Column(scale=3): with gr.Row(): video_stage2 = gr.Video(label='High-Quality Long Video (Preview or Full)', show_label=True, interactive=False, show_download_button=True, height=700) with gr.Row(): run_button_stage2 = gr.Button("Long Video Generation (full high-quality)") inputs_t2v = [prompt_stage1, num_frames, seed, image_stage1] inputs_v2v = [prompt_stage1, num_frames, seed, image_stage1, video_stage1] run_button_stage1.click(fn=generate, inputs=inputs_t2v, outputs=[image_stage1, video_stage1, video_stage2]) run_button_stage2.click(fn=enhance, inputs=inputs_v2v, outputs=[image_stage1, video_stage2]) gr.Examples(examples=examples, inputs=inputs_v2v, outputs=[image_stage1, video_stage2], fn=enhance, cache_examples=True, run_on_click=False, ) ''' ''' gr.HTML("""

Version: v1.0

Caution: We would like the raise the awareness of users of this demo of its potential issues and concerns. Like previous large foundation models, StreamingSVD could be problematic in some cases, partially we use pretrained ModelScope, therefore StreamingSVD can Inherit Its Imperfections. So far, we keep all features available for research testing both to show the great potential of the StreamingSVD framework and to collect important feedback to improve the model in the future. We welcome researchers and users to report issues with the HuggingFace community discussion feature or email the authors.

Biases and content acknowledgement: Beware that StreamingSVD may output content that reinforces or exacerbates societal biases, as well as realistic faces, pornography, and violence. StreamingSVD in this demo is meant only for research purposes.

""") if on_huggingspace: demo.queue(max_size=20) demo.launch(debug=True) else: demo.queue(api_open=False).launch(share=args.public_access)