Source Code/app_text_to_video.py

# ==================================================================================================
# ZERO-SHOT-VIDEO-GENERATION - app_text_to_video.py (Gradio UI Components)
# ==================================================================================================
# 
# 📝 DESCRIPTION
# This module constructs the structural interface for the Text2Video-Zero generation task. It 
# defines the modular Gradio UI components, formulates the layout parameters, and specifies the 
# data bindings between visual controls (like sliders, dropdowns, and buttons) and the underlying 
# neural processing model. Designed for modularity, it manages state interactions specifically for 
# translating textual representations into dynamic video sequences.
#
# 👤 AUTHORS
# - Amey Thakur (https://github.com/Amey-Thakur)
#
# 🤝🏻 CREDITS
# Based directly on the foundational logic of Text2Video-Zero.
# Source Authors: Picsart AI Research (PAIR), UT Austin, U of Oregon, UIUC
# Reference: https://arxiv.org/abs/2303.13439
#
# 🔗 PROJECT LINKS
# Repository: https://github.com/Amey-Thakur/ZERO-SHOT-VIDEO-GENERATION
# Live Demo: https://huggingface.co/spaces/ameythakur/Zero-Shot-Video-Generation
# Video Demo: https://youtu.be/za9hId6UPoY
#
# 📅 RELEASE DATE
# November 22, 2023
#
# 📜 LICENSE
# Released under the MIT License
# ==================================================================================================

import gradio as gr
from model import Model
import os
from hf_utils import get_model_list

# Determine the operational execution context.
on_huggingspace = os.environ.get("SPACE_ID") is not None

# Predefined contextual exemplars establishing baseline structural validation inputs.
# These prompts generate optimal temporal consistency in generated outputs utilizing the
# latent diffusion methodology. Each example must provide values for ALL bound inputs
# (prompt, model_name, video_length) to prevent NoneType errors during example caching.
examples = [
    ["an astronaut waving the arm on the moon", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["a sloth surfing on a wakeboard", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["an astronaut walking on a street", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["a cute cat walking on grass", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["a horse is galloping on a street", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["an astronaut is skiing down the hill", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["a gorilla walking alone down the street", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["a gorilla dancing on times square", "dreamlike-art/dreamlike-photoreal-2.0", 2],
    ["A panda dancing dancing like crazy on Times Square", "dreamlike-art/dreamlike-photoreal-2.0", 2],
]


def create_demo(model: Model):
    """
    Constructs and returns the interactive elements of the Gradio interface for textual inputs.
    Binds the local inference 'model' context to user-facing input handlers to coordinate state 
    between the UI framework and the PyTorch execution context.
    """
    import torch
    is_cpu = not torch.cuda.is_available()

    # Wrapper function ensuring correct keyword argument mapping between the Gradio UI
    # components and the model's process_text2video method. This prevents positional argument
    # misalignment (e.g., video_length being passed as motion_field_strength_x).
    def generate_video(prompt, model_name, video_length):
        return model.process_text2video(
            prompt=prompt,
            model_name=model_name,
            video_length=int(video_length),
        )

    # Instantiate the declarative layout constructor.
    with gr.Blocks() as demo:
        with gr.Row():
            gr.HTML(
                """
                <div style="background: rgba(142,45,226,0.1); padding: 1.5rem; border-left: 5px solid #8E2DE2; border-radius: 10px; margin-bottom: 1.5rem; display: flex; justify-content: space-between; align-items: center; flex-wrap: wrap; gap: 1rem;">
                    <div style="flex: 1; min-width: 300px;">
                        <h2 style="font-weight: 700; font-size: 1.6rem; margin: 0; color: #4A00E0;">
                            Zero-Shot Video Studio
                        </h2>
                        <p style="margin-top: 0.5rem; color: #555; font-size: 1rem; font-weight: 500; line-height: 1.4;">
                            Transform cinematic text prompts into dynamic, temporally consistent AI video. Choose a diffusion model, describe your vision, and generate instantly.
                        </p>
                    </div>
                    <a href="https://colab.research.google.com/github/Amey-Thakur/ZERO-SHOT-VIDEO-GENERATION/blob/main/ZERO-SHOT-VIDEO-GENERATION.ipynb" target="_blank" style="text-decoration: none; transition: all 0.3s cubic-bezier(0.4, 0, 0.2, 1); display: inline-flex; align-items: center; background: white; padding: 0.7rem 1.2rem; border-radius: 12px; border: 1px solid rgba(142,45,226,0.15); box-shadow: 0 4px 15px rgba(0,0,0,0.05);">
                        <div style="display: flex; flex-direction: column; align-items: flex-start; margin-right: 1rem;">
                            <span style="font-size: 0.7rem; font-weight: 800; color: #8E2DE2; text-transform: uppercase; letter-spacing: 1px; margin-bottom: 2px;">Neural Architecture</span>
                            <span style="font-size: 0.95rem; font-weight: 600; color: #4A00E0;">Verified Research Notebook</span>
                        </div>
                        <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Colab" style="height: 22px;"/>
                    </a>
                </div>
                """
            )

        if is_cpu:
            gr.HTML(
                """
                <div style="background: rgba(255,165,0,0.1); padding: 1rem; border-left: 4px solid #FFA500; border-radius: 8px; margin-bottom: 1rem;">
                    <p style="margin: 0; color: #856404; font-size: 0.9rem; line-height: 1.6;">
                        <strong>⚡ CPU Mode.</strong> Running on free-tier hardware. Resolution and frames are reduced to fit this environment.
                        Full resolution on T4 GPU works on Google Colab. Click the notebook link above to try it.
                    </p>
                </div>
                """
            )

        with gr.Row(equal_height=False):
            with gr.Column(scale=1, variant="panel"):
                gr.Markdown("### ✨ Model & Concept Configuration")
                # Configuration block controlling diffusion model weights and textual targets.
                model_name = gr.Dropdown(
                    label="Diffusion Strategy (Model)",
                    choices=get_model_list(),
                    value="dreamlike-art/dreamlike-photoreal-2.0",
                )
                prompt = gr.Textbox(
                    label='Cinematic Prompt', 
                    placeholder="Describe the scene in detail (e.g. 'an astronaut waving the arm on the moon')...",
                    lines=3
                )
                run_button = gr.Button(value='Generate Sequence 🎬', variant='primary', size="lg")
                
                # Expandable execution variables defining trajectory lengths (temporal depth).
                with gr.Accordion('🛠️ Advanced Options', open=False):
                    
                    # Adapting video constraints algorithmically based on the execution domain constraints.
                    if is_cpu:
                        video_length = gr.Slider(
                            label="Video Timeline (Frames)", minimum=2, maximum=4, step=1, value=2)
                    elif on_huggingspace:
                        video_length = gr.Slider(
                            label="Video Timeline (Frames)", minimum=8, maximum=16, step=1, value=8)
                    else:
                        video_length = gr.Number(
                            label="Video Timeline (Frames)", value=8, precision=0)

            with gr.Column(scale=1):
                # Instantiation of the rendering element to visualize synthesized structures.
                gr.Markdown("### 🎞️ Output Stream")
                result = gr.Video(label="Synthesized Video Result", height=380)

        inputs = [
            prompt,
            model_name,
            video_length,
        ]

        # Bind curated input permutations to expedite visualization pathways.
        # cache_examples is disabled: video generation is GPU-intensive and will timeout
        # on free-tier Spaces. Users click "Generate" to run inference on-demand instead.
        gr.Examples(examples=examples,
                    inputs=inputs,
                    outputs=result,
                    fn=generate_video,
                    run_on_click=False,
                    cache_examples=False,
                    )

        # Trigger execution of the generative framework upon interactive activation.
        run_button.click(fn=generate_video,
                         inputs=inputs,
                         outputs=result,)
    return demo