app.py

import os
import subprocess
import sys
from concurrent.futures import ThreadPoolExecutor

# Enable fast downloads
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"

# Disable torch.compile / dynamo before any torch import
os.environ["TORCH_COMPILE_DISABLE"] = "1"
os.environ["TORCHDYNAMO_DISABLE"] = "1"

# Install xformers for memory-efficient attention
subprocess.run([sys.executable, "-m", "pip", "install", "xformers==0.0.32.post2", "--no-build-isolation"], check=False)

# Clone LTX-2 repo at a pinned compatible commit and install packages
LTX_REPO_URL = "https://github.com/Lightricks/LTX-2.git"
LTX_REPO_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "LTX-2")
LTX_COMMIT = "ae855f8538843825f9015a419cf4ba5edaf5eec2"

if os.path.exists(LTX_REPO_DIR):
    print(f"Removing existing repo at {LTX_REPO_DIR}...")
    subprocess.run(["rm", "-rf", LTX_REPO_DIR], check=True)

print(f"Cloning {LTX_REPO_URL}...")
subprocess.run(["git", "clone", LTX_REPO_URL, LTX_REPO_DIR], check=True)

print(f"Checking out commit {LTX_COMMIT}...")
subprocess.run(["git", "-C", LTX_REPO_DIR, "checkout", LTX_COMMIT], check=True)

print("Installing ltx-core and ltx-pipelines from pinned repo commit...")
subprocess.run(
    [
        sys.executable, "-m", "pip", "install",
        "--force-reinstall", "--no-deps",
        "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-core"),
        "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines"),
    ],
    check=True,
)

sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines", "src"))
sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-core", "src"))

import logging
import random
import tempfile
from pathlib import Path

import torch
torch._dynamo.config.suppress_errors = True
torch._dynamo.config.disable = True

# Critical workaround: Replace inference_mode with no_grad
# Avoids "inference tensor" failures in spatial upsampler and VAE decoder
torch.inference_mode = torch.no_grad

import spaces
import gradio as gr
import numpy as np
from huggingface_hub import hf_hub_download, snapshot_download

from ltx_pipelines.distilled import DistilledPipeline
from ltx_pipelines.utils.args import ImageConditioningInput
from ltx_pipelines.utils.media_io import encode_video

# Patch attention backend into the LTX attention module.
import torch.nn.functional as F
from ltx_core.model.transformer import attention as _attn_mod

def _sdpa_as_mea(query, key, value, attn_bias=None, scale=None, **kwargs):
    # xformers memory_efficient_attention: (B, S, H, D) -> (B, S, H, D)
    # torch SDPA:                          (B, H, S, D) -> (B, H, S, D)
    q, k, v = query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2)
    return F.scaled_dot_product_attention(q, k, v, scale=scale).transpose(1, 2)

_cap = torch.cuda.get_device_capability() if torch.cuda.is_available() else (0, 0)
_use_xformers = False
if _cap < (12, 0):
    try:
        from xformers.ops import memory_efficient_attention as _mea
        _attn_mod.memory_efficient_attention = _mea
        _use_xformers = True
        print(f"[ATTN] Using xformers memory_efficient_attention")
    except Exception as e:
        print(f"[ATTN] xformers unavailable ({e}), falling back to SDPA")

if not _use_xformers:
    _attn_mod.memory_efficient_attention = _sdpa_as_mea
    print(f"[ATTN] Using SDPA fallback (sm_{_cap[0]}{_cap[1]})")

logging.getLogger().setLevel(logging.INFO)

MAX_SEED = np.iinfo(np.int32).max
DEFAULT_PROMPT = (
    "An astronaut hatches from a fragile egg on the surface of the Moon, "
    "the shell cracking and peeling apart in gentle low-gravity motion. "
    "Fine lunar dust lifts and drifts outward with each movement, floating "
    "in slow arcs before settling back onto the ground."
)
DEFAULT_FRAME_RATE = 24.0

# Resolution presets: (width, height)
RESOLUTIONS = {
    "high": {"16:9": (1536, 1024), "9:16": (1024, 1536), "1:1": (1024, 1024)},
    "low": {"16:9": (512, 320), "9:16": (320, 512), "1:1": (512, 512)},
}


# Model repos
CHECKPOINT_REPO = "SulphurAI/Sulphur-2-base"
LTX_MODEL_REPO = "Lightricks/LTX-2.3"
GEMMA_REPO = "Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"

# Download model checkpoints in parallel for speed
print("=" * 80)
print("Downloading Element-16 (pre-distilled) + Gemma (parallel)...")
print("=" * 80)

def download_checkpoint():
    # Use pre-distilled checkpoint - no LoRA needed
    return hf_hub_download(repo_id=CHECKPOINT_REPO, filename="sulphur_distil_bf16.safetensors")

def download_upsampler():
    return hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")

def download_gemma():
    return snapshot_download(repo_id=GEMMA_REPO)

with ThreadPoolExecutor(max_workers=3) as executor:
    future_checkpoint = executor.submit(download_checkpoint)
    future_upsampler = executor.submit(download_upsampler)
    future_gemma = executor.submit(download_gemma)

    checkpoint_path = future_checkpoint.result()
    spatial_upsampler_path = future_upsampler.result()
    gemma_root = future_gemma.result()

print(f"Checkpoint: {checkpoint_path}")
print(f"Spatial upsampler: {spatial_upsampler_path}")
print(f"Gemma root: {gemma_root}")

# Initialize pipeline with pre-distilled checkpoint (no LoRA needed)
pipeline = DistilledPipeline(
    distilled_checkpoint_path=checkpoint_path,
    spatial_upsampler_path=spatial_upsampler_path,
    gemma_root=gemma_root,
    loras=(),
)

# Preload all models for ZeroGPU tensor packing
print("Preloading all pipeline components via model_ledger...")

# DistilledPipeline uses model_ledger similar to other pipelines
ledger = pipeline.model_ledger
_transformer = ledger.transformer()
_video_encoder = ledger.video_encoder()
_video_decoder = ledger.video_decoder()
_spatial_upsampler = ledger.spatial_upsampler()
_text_encoder = ledger.text_encoder()
_embeddings_processor = ledger.gemma_embeddings_processor()
_audio_encoder = ledger.audio_encoder()
_audio_decoder = ledger.audio_decoder()
_vocoder = ledger.vocoder()

# Replace ledger methods with lambdas returning preloaded instances
ledger.transformer = lambda: _transformer
ledger.video_encoder = lambda: _video_encoder
ledger.video_decoder = lambda: _video_decoder
ledger.spatial_upsampler = lambda: _spatial_upsampler
ledger.text_encoder = lambda: _text_encoder
ledger.gemma_embeddings_processor = lambda: _embeddings_processor
ledger.audio_encoder = lambda: _audio_encoder
ledger.audio_decoder = lambda: _audio_decoder
ledger.vocoder = lambda: _vocoder

print("All models preloaded!")

print("=" * 80)
print("Pipeline ready!")
print("=" * 80)


def log_memory(tag: str):
    if torch.cuda.is_available():
        allocated = torch.cuda.memory_allocated() / 1024**3
        peak = torch.cuda.max_memory_allocated() / 1024**3
        free, total = torch.cuda.mem_get_info()
        print(f"[VRAM {tag}] allocated={allocated:.2f}GB peak={peak:.2f}GB free={free / 1024**3:.2f}GB total={total / 1024**3:.2f}GB")


def detect_aspect_ratio(image) -> str:
    if image is None:
        return "16:9"
    if hasattr(image, "size"):
        w, h = image.size
    elif hasattr(image, "shape"):
        h, w = image.shape[:2]
    else:
        return "16:9"
    ratio = w / h
    candidates = {"16:9": 16 / 9, "9:16": 9 / 16, "1:1": 1.0}
    return min(candidates, key=lambda k: abs(ratio - candidates[k]))


def on_image_upload(first_image, last_image, high_res):
    ref_image = first_image if first_image is not None else last_image
    aspect = detect_aspect_ratio(ref_image)
    tier = "high" if high_res else "low"
    w, h = RESOLUTIONS[tier][aspect]
    return gr.update(value=w), gr.update(value=h)


def on_highres_toggle(first_image, last_image, high_res):
    ref_image = first_image if first_image is not None else last_image
    aspect = detect_aspect_ratio(ref_image)
    tier = "high" if high_res else "low"
    w, h = RESOLUTIONS[tier][aspect]
    return gr.update(value=w), gr.update(value=h)


DEFAULT_NEGATIVE_PROMPT = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走, blurry, glasses, deformed, subtitles, text, captions, worst quality, low quality, inconsistent motion, jittery, distorted"


def remove_music_demucs(input_video_path: str, output_video_path: str) -> bool:
    """Remove background music from video using Demucs, keeping only vocals."""
    import subprocess
    import tempfile
    from pathlib import Path

    try:
        with tempfile.TemporaryDirectory() as tmpdir:
            tmpdir = Path(tmpdir)

            # Extract audio from video
            audio_in = tmpdir / "audio.wav"
            extract_cmd = [
                'ffmpeg', '-y', '-i', input_video_path,
                '-vn', '-acodec', 'pcm_s16le', '-ar', '44100', '-ac', '2',
                str(audio_in)
            ]
            result = subprocess.run(extract_cmd, capture_output=True, text=True)
            if result.returncode != 0:
                print(f"[demucs] Failed to extract audio: {result.stderr[-200:]}")
                return False

            print(f"[demucs] Running music separation...")

            import soundfile as sf
            from demucs.pretrained import get_model
            from demucs.apply import apply_model

            # Load model (cached after first run)
            model = get_model('htdemucs')
            model.to('cuda')
            model.eval()

            # Load audio
            data, sr = sf.read(str(audio_in))
            wav = torch.from_numpy(data.T).float()
            if wav.dim() == 1:
                wav = wav.unsqueeze(0)

            # Resample if needed
            if sr != model.samplerate:
                import torchaudio
                wav = torchaudio.functional.resample(wav, sr, model.samplerate)

            wav = wav.unsqueeze(0).to('cuda')

            # Separate sources
            with torch.no_grad():
                sources = apply_model(model, wav, overlap=0.25, progress=False)

            # Keep only vocals (index 3)
            vocals = sources[0, 3].cpu()

            # Save vocals
            audio_out = tmpdir / "vocals.wav"
            audio_np = vocals.numpy().T
            sf.write(str(audio_out), audio_np, model.samplerate)

            print(f"[demucs] Merging vocals back with video...")
            merge_cmd = [
                'ffmpeg', '-y',
                '-i', input_video_path,
                '-i', str(audio_out),
                '-c:v', 'copy',
                '-map', '0:v:0', '-map', '1:a:0',
                '-c:a', 'aac', '-b:a', '128k',
                '-shortest',
                output_video_path
            ]
            result = subprocess.run(merge_cmd, capture_output=True, text=True)
            if result.returncode != 0:
                print(f"[demucs] Failed to merge: {result.stderr[-200:]}")
                return False

            print(f"[demucs] Successfully removed music")
            return True

    except Exception as e:
        print(f"[demucs] Error: {e}")
        import traceback
        traceback.print_exc()
        return False


def apply_gaussian_blur(video_tensor: torch.Tensor, blur_amount: int) -> torch.Tensor:
    """Apply Gaussian blur to video tensor. Video shape: [frames, H, W, C]"""
    if blur_amount <= 0:
        return video_tensor

    from torchvision.transforms.functional import gaussian_blur

    # Ensure kernel size is odd and at least 3
    kernel_size = blur_amount * 2 + 1
    sigma = blur_amount / 2.0

    # Video tensor is [frames, H, W, C], but gaussian_blur expects [batch, C, H, W]
    # Permute to [frames, C, H, W]
    video_tensor = video_tensor.permute(0, 3, 1, 2)

    blurred = gaussian_blur(video_tensor, kernel_size=[kernel_size, kernel_size], sigma=[sigma, sigma])

    # Permute back to [frames, H, W, C]
    blurred = blurred.permute(0, 2, 3, 1)

    return blurred


@spaces.GPU(duration=90)
@torch.inference_mode()
def generate_video(
    first_image,
    last_image,
    prompt: str,
    duration: float,
    enhance_prompt: bool = True,
    seed: int = 42,
    randomize_seed: bool = True,
    height: int = 320,
    width: int = 512,
    negative_prompt: str = DEFAULT_NEGATIVE_PROMPT,
    blur_amount: int = 0,
    remove_music: bool = False,
    progress=gr.Progress(track_tqdm=True),
):
    try:
        torch.cuda.reset_peak_memory_stats()
        log_memory("start")

        current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)

        frame_rate = DEFAULT_FRAME_RATE
        num_frames = int(duration * frame_rate) + 1
        num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1

        print(f"Generating: {height}x{width}, {num_frames} frames ({duration}s), seed={current_seed}")

        images = []
        output_dir = Path("outputs")
        output_dir.mkdir(exist_ok=True)

        if first_image is not None:
            temp_first_path = output_dir / f"temp_first_{current_seed}.jpg"
            if hasattr(first_image, "save"):
                first_image.save(temp_first_path)
            else:
                temp_first_path = Path(first_image)
            images.append(ImageConditioningInput(path=str(temp_first_path), frame_idx=0, strength=1.0))

        if last_image is not None:
            temp_last_path = output_dir / f"temp_last_{current_seed}.jpg"
            if hasattr(last_image, "save"):
                last_image.save(temp_last_path)
            else:
                temp_last_path = Path(last_image)
            images.append(ImageConditioningInput(path=str(temp_last_path), frame_idx=num_frames - 1, strength=1.0))

        from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number

        tiling_config = TilingConfig.default()
        video_chunks_number = get_video_chunks_number(num_frames, tiling_config)

        log_memory("before pipeline call")

        # Run inference - DistilledPipeline has simpler API
        video_frames_iter, audio = pipeline(
            prompt=prompt,
            seed=current_seed,
            height=int(height),
            width=int(width),
            num_frames=num_frames,
            frame_rate=frame_rate,
            images=images,
            enhance_prompt=enhance_prompt,
        )

        # Collect video frames
        frames = [frame for frame in video_frames_iter]
        video_tensor = torch.cat(frames, dim=0) if len(frames) > 1 else frames[0]

        log_memory("after pipeline call")

        # Apply Gaussian blur if requested (for censoring/teaser effect)
        if blur_amount > 0:
            print(f"Applying Gaussian blur (amount={blur_amount})...")
            video_tensor = apply_gaussian_blur(video_tensor, blur_amount)
            log_memory("after blur")

        output_path = tempfile.mktemp(suffix=".mp4")
        encode_video(
            video=video_tensor,
            fps=frame_rate,
            audio=audio,
            output_path=output_path,
            video_chunks_number=video_chunks_number,
        )

        log_memory("after encode_video")

        # Remove background music if requested
        if remove_music:
            print(f"Removing background music with Demucs...")
            processed_path = tempfile.mktemp(suffix=".mp4")
            success = remove_music_demucs(output_path, processed_path)
            if success:
                output_path = processed_path
                log_memory("after demucs")
            else:
                print(f"Warning: Music removal failed, using original video")

        return str(output_path), current_seed

    except Exception as e:
        import traceback
        log_memory("on error")
        print(f"Error: {str(e)}\n{traceback.format_exc()}")
        return None, current_seed


with gr.Blocks(title="Element-16 Video", delete_cache=(3600, 7200)) as demo:  # cleanup: check every 1h, delete files >2h old
    gr.Markdown("# Element-16: Fast Video Generation with Frame Conditioning")
    gr.Markdown(
        "High quality video + audio generation with first and last frame conditioning. "
        "Optimized fp8 model for faster inference. "
        "[[code]](https://github.com/Lightricks/LTX-2)"
    )

    with gr.Row():
        with gr.Column():
            with gr.Row():
                first_image = gr.Image(label="First Frame (Optional)", type="pil")
                last_image = gr.Image(label="Last Frame (Optional)", type="pil")
            prompt = gr.Textbox(
                label="Prompt",
                info="for best results - make it as elaborate as possible",
                value="Make this image come alive with cinematic motion, smooth animation",
                lines=3,
                placeholder="Describe the motion and animation you want...",
            )
            duration = gr.Slider(label="Duration (seconds)", minimum=1.0, maximum=10.0, value=3.0, step=0.1)

            generate_btn = gr.Button("Generate Video", variant="primary", size="lg")

            with gr.Accordion("Advanced Settings", open=False):
                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, value=10, step=1)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                with gr.Row():
                    width = gr.Number(label="Width", value=512, precision=0)
                    height = gr.Number(label="Height", value=320, precision=0)
                with gr.Row():
                    enhance_prompt = gr.Checkbox(label="Enhance Prompt", value=False)
                    high_res = gr.Checkbox(label="High Resolution", value=False)
                with gr.Row():
                    blur_amount = gr.Number(label="Blur (0=off, 36=heavy)", value=0, precision=0)
                    remove_music = gr.Checkbox(label="Remove Music", value=False)
                negative_prompt = gr.Textbox(
                    label="Negative Prompt",
                    value=DEFAULT_NEGATIVE_PROMPT,
                    lines=3,
                    placeholder="What to avoid in the generated video...",
                )

        with gr.Column():
            output_video = gr.Video(label="Generated Video", autoplay=True)

    gr.Examples(
        examples=[
            [
                None,
                "pinkknit.jpg",
                "The camera falls downward through darkness as if dropped into a tunnel. "
                "As it slows, five friends wearing pink knitted hats and sunglasses lean "
                "over and look down toward the camera with curious expressions. The lens "
                "has a strong fisheye effect, creating a circular frame around them. They "
                "crowd together closely, forming a symmetrical cluster while staring "
                "directly into the lens.",
                3.0,
                False,
                42,
                True,
                1024,
                1024,
            ],
        ],
        inputs=[
            first_image, last_image, prompt, duration,
            enhance_prompt, seed, randomize_seed, height, width,
        ],
    )

    first_image.change(
        fn=on_image_upload,
        inputs=[first_image, last_image, high_res],
        outputs=[width, height],
    )

    last_image.change(
        fn=on_image_upload,
        inputs=[first_image, last_image, high_res],
        outputs=[width, height],
    )

    high_res.change(
        fn=on_highres_toggle,
        inputs=[first_image, last_image, high_res],
        outputs=[width, height],
    )

    generate_btn.click(
        fn=generate_video,
        inputs=[
            first_image, last_image, prompt, duration, enhance_prompt,
            seed, randomize_seed, height, width, negative_prompt, blur_amount, remove_music,
        ],
        outputs=[output_video, seed],
    )


css = """
.fillable{max-width: 1200px !important}
"""

if __name__ == "__main__":
    demo.launch(theme=gr.themes.Citrus(), css=css)