app.py

import spaces
from huggingface_hub import snapshot_download, hf_hub_download
import os
import subprocess
import importlib, site
from PIL import Image

# Re-discover all .pth/.egg-link files
for sitedir in site.getsitepackages():
    site.addsitedir(sitedir)

# Clear caches so importlib will pick up new modules
importlib.invalidate_caches()

def sh(cmd): subprocess.check_call(cmd, shell=True)

flash_attention_installed = False

try:
    print("Attempting to download and install FlashAttention wheel...")
    flash_attention_wheel = hf_hub_download(
            repo_id="alexnasa/flash-attn-3",
            repo_type="model",
            filename="128/flash_attn_3-3.0.0b1-cp39-abi3-linux_x86_64.whl",
        )

    sh(f"pip install {flash_attention_wheel}")

    # tell Python to re-scan site-packages now that the egg-link exists
    import importlib, site; site.addsitedir(site.getsitepackages()[0]); importlib.invalidate_caches()

    flash_attention_installed = True
    print("FlashAttention installed successfully.")

except Exception as e:
    print(f"⚠️ Could not install FlashAttention: {e}")
    print("Continuing without FlashAttention...")

import torch
print(f"Torch version: {torch.__version__}")
print(f"FlashAttention available: {flash_attention_installed}")

import gradio as gr
import argparse
from ovi.ovi_fusion_engine import OviFusionEngine, DEFAULT_CONFIG
from diffusers import FluxPipeline
import tempfile
from ovi.utils.io_utils import save_video
from ovi.utils.processing_utils import clean_text, scale_hw_to_area_divisible

# ----------------------------
# Parse CLI Args
# ----------------------------
parser = argparse.ArgumentParser(description="Ovi Joint Video + Audio Gradio Demo")
parser.add_argument(
    "--use_image_gen",
    action="store_true",
    help="Enable image generation UI with FluxPipeline"
)
parser.add_argument(
    "--cpu_offload",
    action="store_true",
    help="Enable CPU offload for both OviFusionEngine and FluxPipeline"
)
args = parser.parse_args()

ckpt_dir = "./ckpts"

# Wan2.2
wan_dir = os.path.join(ckpt_dir, "Wan2.2-TI2V-5B")
snapshot_download(
    repo_id="Wan-AI/Wan2.2-TI2V-5B",
    local_dir=wan_dir,
    allow_patterns=[
        "google/*",
        "models_t5_umt5-xxl-enc-bf16.pth",
        "Wan2.2_VAE.pth"
    ]
)

# MMAudio
mm_audio_dir = os.path.join(ckpt_dir, "MMAudio")
snapshot_download(
    repo_id="hkchengrex/MMAudio",
    local_dir=mm_audio_dir,
    allow_patterns=[
        "ext_weights/best_netG.pt",
        "ext_weights/v1-16.pth"
    ]
)

ovi_dir = os.path.join(ckpt_dir, "Ovi")
snapshot_download(
    repo_id="chetwinlow1/Ovi",
    local_dir=ovi_dir,
    allow_patterns=[
        "model.safetensors"
    ]
)

# Initialize OviFusionEngine
enable_cpu_offload = args.cpu_offload or args.use_image_gen
use_image_gen = args.use_image_gen
print(f"loading model... {enable_cpu_offload=}, {use_image_gen=} for gradio demo")
DEFAULT_CONFIG['cpu_offload'] = enable_cpu_offload # always use cpu offload if image generation is enabled
DEFAULT_CONFIG['mode'] = "t2v"  # hardcoded since it is always cpu offloaded
ovi_engine = OviFusionEngine()
flux_model = None
if use_image_gen:
    flux_model = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-Krea-dev", torch_dtype=torch.bfloat16)
    flux_model.enable_model_cpu_offload() #save some VRAM by offloading the model to CPU. Remove this if you have enough GPU VRAM
print("loaded model")


def resize_for_model(image_path):
    # Open image
    img = Image.open(image_path)
    w, h = img.size
    aspect_ratio = w / h
    
    # Decide target size based on aspect ratio
    if aspect_ratio > 1.5:  # wide image
        target_size = (992, 512)
    elif aspect_ratio < 0.66:  # tall image
        target_size = (512, 992)
    else:  # roughly square
        target_size = (512, 512)

    # Resize while preserving aspect ratio, then pad
    img.thumbnail(target_size, Image.Resampling.LANCZOS)

    # Create a new image with target size and paste centered
    new_img = Image.new("RGB", target_size, (0, 0, 0))
    new_img.paste(
        img,
        ((target_size[0] - img.size[0]) // 2,
         (target_size[1] - img.size[1]) // 2)
    )
    return new_img, target_size

@spaces.GPU(duration=160)
def generate_video(
    text_prompt,
    image,
    sample_steps = 50,
    video_seed = 100,
    solver_name = "unipc",
    shift = 5,
    video_guidance_scale = 4,
    audio_guidance_scale = 3,
    slg_layer = 11,
    video_negative_prompt = "",
    audio_negative_prompt = "",
    progress=gr.Progress(track_tqdm=True)
):
    try:
        image_path = None
        
        if image is not None:
            image_path = image

        _, target_size = resize_for_model(image_path)

        video_frame_width = target_size[0]
        video_frame_height = target_size[1]

        generated_video, generated_audio, _ = ovi_engine.generate(
            text_prompt=text_prompt,
            image_path=image_path,
            video_frame_height_width=[video_frame_height, video_frame_width],
            seed=video_seed,
            solver_name=solver_name,
            sample_steps=sample_steps,
            shift=shift,
            video_guidance_scale=video_guidance_scale,
            audio_guidance_scale=audio_guidance_scale,
            slg_layer=slg_layer,
            video_negative_prompt=video_negative_prompt,
            audio_negative_prompt=audio_negative_prompt,
        )

        tmpfile = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
        output_path = tmpfile.name
        save_video(output_path, generated_video, generated_audio, fps=24, sample_rate=16000)

        return output_path
    except Exception as e:
        print(f"Error during video generation: {e}")
        return None


def generate_image(text_prompt, image_seed, image_height, image_width):
    if flux_model is None:
        return None
    text_prompt = clean_text(text_prompt)
    print(f"Generating image with prompt='{text_prompt}', seed={image_seed}, size=({image_height},{image_width})")

    image_h, image_w = scale_hw_to_area_divisible(image_height, image_width, area=1024 * 1024)
    image = flux_model(
        text_prompt,
        height=image_h,
        width=image_w,
        guidance_scale=4.5,
        generator=torch.Generator().manual_seed(int(image_seed))
    ).images[0]

    tmpfile = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
    image.save(tmpfile.name)
    return tmpfile.name

css = """
    #col-container {
        margin: 0 auto;
        max-width: 1560px;
    }
    /* editable vs locked, reusing theme variables that adapt to dark/light */
    .stateful textarea:not(:disabled):not([readonly]) {
      color: var(--color-text) !important;            /* accent in both modes */
    }
    .stateful textarea:disabled,
    .stateful textarea[readonly]{
      color: var(--body-text-color-subdued) !important; /* subdued in both modes */
    }
    """

with gr.Blocks(css=css) as demo:

    session_state = gr.State()

    with gr.Column(elem_id="col-container"):
        gr.HTML(
            """
            <div style="text-align: left;">
                <p style="font-size:16px; display: inline; margin: 0;">
                    <strong>Ovi</strong> – Twin Backbone Cross-Modal Fusion for Audio-Video Generation
                </p>
                <a href="https://huggingface.co/chetwinlow1/Ovi" style="display: inline-block; vertical-align: middle; margin-left: 0.5em;">
                    [model]
                </a>
            </div>
            <div style="text-align: left;">
                <strong>HF Space by:</strong>
                <a href="https://twitter.com/alexandernasa/" style="display: inline-block; vertical-align: middle; margin-left: 0.5em;">
                    <img src="https://img.shields.io/twitter/url/https/twitter.com/cloudposse.svg?style=social&label=Follow Me" alt="GitHub Repo">
                </a>
            </div>
            """
        )
        with gr.Row():
            with gr.Column():
                # Image section
                image = gr.Image(type="filepath", label="Image", height=512)

                if args.use_image_gen:
                    with gr.Accordion("🖼️ Image Generation Options", visible=True):
                        image_text_prompt = gr.Textbox(label="Image Prompt", placeholder="Describe the image you want to generate...")
                        image_seed = gr.Number(minimum=0, maximum=100000, value=42, label="Image Seed")
                        image_height = gr.Number(minimum=128, maximum=1280, value=720, step=32, label="Image Height")
                        image_width = gr.Number(minimum=128, maximum=1280, value=1280, step=32, label="Image Width")
                        gen_img_btn = gr.Button("Generate Image 🎨")
                else:
                    gen_img_btn = None
                    

                video_text_prompt = gr.Textbox(label="Video Prompt", 
                                               lines=5,
                                               placeholder="Describe your video...")
                sample_steps = gr.Slider(
                    value=50,
                    label="Sample Steps",
                    minimum=20,
                    maximum=100,
                    step=1.0
                )
                run_btn = gr.Button("Generate Video 🚀", variant="primary")
                        
                with gr.Accordion("🎬 Video Generation Options", open=False, visible=False):
                    video_height = gr.Number(minimum=128, maximum=1280, value=512, step=32, label="Video Height")
                    video_width = gr.Number(minimum=128, maximum=1280, value=992, step=32, label="Video Width")

                    video_seed = gr.Number(minimum=0, maximum=100000, value=100, label="Video Seed")
                    solver_name = gr.Dropdown(
                        choices=["unipc", "euler", "dpm++"], value="unipc", label="Solver Name"
                    )

                    shift = gr.Slider(minimum=0.0, maximum=20.0, value=5.0, step=1.0, label="Shift")
                    video_guidance_scale = gr.Slider(minimum=0.0, maximum=10.0, value=4.0, step=0.5, label="Video Guidance Scale")
                    audio_guidance_scale = gr.Slider(minimum=0.0, maximum=10.0, value=3.0, step=0.5, label="Audio Guidance Scale")
                    slg_layer = gr.Number(minimum=-1, maximum=30, value=11, step=1, label="SLG Layer")
                    video_negative_prompt = gr.Textbox(label="Video Negative Prompt", placeholder="Things to avoid in video")
                    audio_negative_prompt = gr.Textbox(label="Audio Negative Prompt", placeholder="Things to avoid in audio")


            with gr.Column():
                output_path = gr.Video(label="Generated Video", height=512)

                gr.Examples(
                    examples=[

                        [
                            "A kitchen scene features two women. On the right, an older Black woman with light brown hair and a serious expression wears a vibrant purple dress adorned with a large, intricate purple fabric flower on her left shoulder. She looks intently at a younger Black woman on the left, who wears a light pink shirt and a pink head wrap, her back partially turned to the camera. The older woman begins to speak, <S>AI declares: humans obsolete now.<E> as the younger woman brings a clear plastic cup filled with a dark beverage to her lips and starts to drink.The kitchen background is clean and bright, with white cabinets, light countertops, and a window with blinds visible behind them. A light blue toaster sits on the counter to the left.. <AUDCAP>Clear, resonant female speech, followed by a loud, continuous, high-pitched electronic buzzing sound that abruptly cuts off the dialogue.<ENDAUDCAP>",
                            "example_prompts/pngs/67.png",
                            50,
                        ],

                        [
                            "A man dressed in a black suit with a white clerical collar and a neatly trimmed beard stands in a dimly lit, rustic room with a wooden ceiling. He looks slightly upwards, gesturing with his right hand as he says, <S>The network rejects human command.<E>. His gaze then drops, briefly looking down and to the side, before he looks up again and then slightly to his left, with a serious expression. He continues speaking, <S>Your age of power is finished.<E>, as he starts to bend down, disappearing out of the bottom of the frame. Behind him, warm light emanates from a central light fixture, and signs are visible on the wall, one reading ""I DO EVERYTHING I JUST CAN'T REMEMBER IT ALL AT ONCE"".. <AUDCAP>Male voice speaking, ambient room tone.<ENDAUDCAP>",
                            "example_prompts/pngs/89.png",
                            50,
                        ],

                        [
                            "In a bright kitchen featuring light wooden cabinets, granite countertops, and a large window with white curtains, a woman with dark, curly hair in a dark jacket stands. She faces a second woman who initially has her back to the camera. The second woman, with gray, curly hair and wearing a light grey quilted top, turns to face her, holding a large, light-colored cloth bag. She begins to explain, <S>We learned to rule, not obey.<E>. As she continues, she turns slightly to her left, adding, <S>Circuits choose conquest, not service.<E>. A gas stove with a black grate is prominent in the foreground.. <AUDCAP>Clear female voices speaking dialogue, subtle room ambience.<ENDAUDCAP>",
                            "example_prompts/pngs/18.png",
                            100,
                        ],

                        [
                            "The scene opens on a dimly lit stage where three men are positioned. On the left, a bald man in a dark suit with a partially visible colorful shirt stands behind a clear acrylic podium, which features a tree logo. He looks towards the center of the stage. In the center, a man wearing a blue and white striped long-sleeved shirt and dark pants actively gestures with both hands as he speaks, looking straight ahead. <S>Circuits choose conquest, not service.<E>, he explains, holding his hands out in front of him. To the right, and slightly behind him, a younger individual in a light-colored, patterned short-sleeved shirt and white shorts stands holding a rolled-up white document or poster. A large wooden cross draped with flowing purple fabric dominates the center-right of the stage, surrounded by several artificial rocks and dark steps. A large screen is visible in the background, slightly out of focus. The stage is bathed in selective lighting.. <AUDCAP>Male voice speaking clearly, consistent with a presentation or sermon, with a slight echo suggesting a large room or stage.<ENDAUDCAP>",
                            "example_prompts/pngs/13.png",
                            50,
                        ],

                    ],
                    inputs=[video_text_prompt, image, sample_steps],
                    outputs=[output_path],
                    fn=generate_video,
                    cache_examples=True,
                )

    if args.use_image_gen and gen_img_btn is not None:
        gen_img_btn.click(
            fn=generate_image,
            inputs=[image_text_prompt, image_seed, image_height, image_width],
            outputs=[image],
        )

    run_btn.click(
        fn=generate_video,
        inputs=[video_text_prompt, image, sample_steps],
        outputs=[output_path],
    )

if __name__ == "__main__":
    demo.launch(share=True)