Create app.py

app.py

@@ -0,0 +1,214 @@
+import os
+import subprocess
+import sys
+
+# Disable torch.compile / dynamo before any torch import
+os.environ["TORCH_COMPILE_DISABLE"] = "1"
+os.environ["TORCHDYNAMO_DISABLE"] = "1"
+
+# Clone LTX-2 repo 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")
+
+if not os.path.exists(LTX_REPO_DIR):
+    print(f"Cloning {LTX_REPO_URL}...")
+    subprocess.run(["git", "clone", "--depth", "1", LTX_REPO_URL, LTX_REPO_DIR], check=True)
+
+# Install ltx-core and ltx-pipelines if not already installed
+try:
+    import ltx_pipelines  # noqa: F401
+except ImportError:
+    print("Installing ltx-core and ltx-pipelines...")
+    subprocess.run(
+        [sys.executable, "-m", "pip", "install", "-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
+
+import torch
+torch._dynamo.config.suppress_errors = True
+torch._dynamo.config.disable = True
+
+import spaces
+import gradio as gr
+import numpy as np
+from huggingface_hub import hf_hub_download, snapshot_download
+
+from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number
+from ltx_core.quantization import QuantizationPolicy
+from ltx_pipelines.distilled import DistilledPipeline
+from ltx_pipelines.utils.args import ImageConditioningInput
+from ltx_pipelines.utils.media_io import encode_video
+
+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."
+)
+DEFAULT_HEIGHT = 1024
+DEFAULT_WIDTH = 1536
+DEFAULT_FRAME_RATE = 24.0
+
+# Download models from Hugging Face
+LTX_MODEL_REPO = "diffusers-internal-dev/ltx-23"
+GEMMA_MODEL_REPO = "google/gemma-3-12b-it-qat-q4_0-unquantized"
+
+print("=" * 80)
+print("Downloading models from Hugging Face...")
+print("=" * 80)
+
+DISTILLED_CHECKPOINT = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-22b-distilled.safetensors")
+SPATIAL_UPSAMPLER = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")
+GEMMA_ROOT = snapshot_download(repo_id=GEMMA_MODEL_REPO)
+
+print(f"Distilled checkpoint: {DISTILLED_CHECKPOINT}")
+print(f"Spatial upsampler: {SPATIAL_UPSAMPLER}")
+print(f"Gemma root: {GEMMA_ROOT}")
+
+# Initialize pipeline
+print("=" * 80)
+print("Loading LTX-2.3 Distilled pipeline...")
+print("=" * 80)
+
+pipeline = DistilledPipeline(
+    distilled_checkpoint_path=DISTILLED_CHECKPOINT,
+    spatial_upsampler_path=SPATIAL_UPSAMPLER,
+    gemma_root=GEMMA_ROOT,
+    loras=[],
+    quantization=QuantizationPolicy.fp8_cast(),
+)
+
+# Preload all models so first request is fast.
+# On ZeroGPU, .to('cuda') is intercepted and actual GPU allocation
+# happens inside the @spaces.GPU decorated function.
+print("Preloading models...")
+ledger = pipeline.model_ledger
+_text_encoder = ledger.text_encoder()
+_transformer = ledger.transformer()
+_video_encoder = ledger.video_encoder()
+_video_decoder = ledger.video_decoder()
+_audio_decoder = ledger.audio_decoder()
+_vocoder = ledger.vocoder()
+_spatial_upsampler = ledger.spatial_upsampler()
+
+ledger.text_encoder = lambda: _text_encoder
+ledger.transformer = lambda: _transformer
+ledger.video_encoder = lambda: _video_encoder
+ledger.video_decoder = lambda: _video_decoder
+ledger.audio_decoder = lambda: _audio_decoder
+ledger.vocoder = lambda: _vocoder
+ledger.spatial_upsampler = lambda: _spatial_upsampler
+
+print("All models preloaded!")
+
+
+@spaces.GPU(duration=300)
+@torch.inference_mode()
+def generate_video(
+    input_image,
+    prompt: str,
+    duration: float,
+    enhance_prompt: bool,
+    seed: int,
+    randomize_seed: bool,
+    height: int,
+    width: int,
+    progress=gr.Progress(track_tqdm=True),
+):
+    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+    num_frames = int(duration * DEFAULT_FRAME_RATE) + 1
+    num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1
+
+    images = []
+    if input_image is not None:
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+            temp_path = f.name
+            if hasattr(input_image, "save"):
+                input_image.save(temp_path)
+            else:
+                from shutil import copy2
+                copy2(str(input_image), temp_path)
+            images = [ImageConditioningInput(path=temp_path, frame_idx=0, strength=1.0)]
+
+    tiling_config = TilingConfig.default()
+    video_chunks_number = get_video_chunks_number(num_frames, tiling_config)
+
+    video, audio = pipeline(
+        prompt=prompt,
+        seed=current_seed,
+        height=int(height),
+        width=int(width),
+        num_frames=num_frames,
+        frame_rate=DEFAULT_FRAME_RATE,
+        images=images,
+        tiling_config=tiling_config,
+        enhance_prompt=enhance_prompt,
+    )
+
+    output_path = tempfile.mktemp(suffix=".mp4")
+    encode_video(
+        video=video,
+        fps=DEFAULT_FRAME_RATE,
+        audio=audio,
+        output_path=output_path,
+        video_chunks_number=video_chunks_number,
+    )
+
+    return output_path, current_seed
+
+
+with gr.Blocks(title="LTX-2.3 Distilled") as demo:
+    gr.Markdown("# LTX-2.3 Distilled (22B): Fast Audio-Video Generation")
+    gr.Markdown(
+        "Fast video + audio generation using the distilled model (8 steps stage 1, 4 steps stage 2). "
+        "[[model]](https://huggingface.co/Lightricks/LTX-2) "
+        "[[code]](https://github.com/Lightricks/LTX-2)"
+    )
+
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(label="Input Image (Optional)", type="pil")
+            prompt = gr.Textbox(
+                label="Prompt",
+                value=DEFAULT_PROMPT,
+                lines=3,
+                placeholder="Describe the video you want to generate...",
+            )
+            with gr.Row():
+                duration = gr.Slider(label="Duration (seconds)", minimum=1.0, maximum=10.0, value=5.0, step=0.5)
+                enhance_prompt = gr.Checkbox(label="Enhance Prompt", value=True)
+
+            generate_btn = gr.Button("Generate Video", variant="primary")
+
+            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=DEFAULT_WIDTH, precision=0)
+                    height = gr.Number(label="Height", value=DEFAULT_HEIGHT, precision=0)
+
+        with gr.Column():
+            output_video = gr.Video(label="Generated Video", autoplay=True)
+
+    generate_btn.click(
+        fn=generate_video,
+        inputs=[
+            input_image, prompt, duration, enhance_prompt,
+            seed, randomize_seed, height, width,
+        ],
+        outputs=[output_video, seed],
+    )
+
+
+if __name__ == "__main__":
+    demo.launch(share=True)