🎬 DeepRat LTX Video - AI Video Generation

Advanced AI-powered video generation using LTX Video model

🚀 Try Demo | 📖 Documentation | 🎨 Examples

---

boxing_cat.mp4

---

✨ Features

• 🎬 Text-to-Video: Generate videos from text descriptions
• 🖼️ Image-to-Video: Animate static images with AI
• 🎞️ Video-to-Video: Transform and enhance existing videos
• 🎯 Multi-Conditioning: Apply multiple conditions at specific frames
• ⚡ High Performance: Optimized for CUDA with CPU fallback
• 🎨 Flexible Control: Fine-tune every aspect of generation

---

📸 Examples

Text-to-Video (PtV)

Entrada	Salida
Example 1 — Majestic Black Lion Seed: 1363812591 Prompt: A beautiful and powerful black lion	1363812591.mp4
Example 2 — Snow-Capped Mountains Seed: 3804031196 Prompt: A view from above of beautiful snow-capped mountains	3804031196.mp4
Example 3 — Tiger Attacking Wild Boar Seed: 1397763684 Prompt: A tiger jumping/attacking a wild boar	1397763684.mp4

---

Image-to-Video (ItV)

Entrada (Imagen)	Salida (Video)
Seed: 1830526882 Prompt: the two cats in the image are fighting each other with kicks and Muay Thai fists in a very active and dizzying way like an action fight	1830526882.mp4
Seed: 738317591 Prompt: The cats from the picture are boxing each other agresivelly	738317591.mp4
Seed: 3858595085 Prompt: take the lion from the drawing and remove the background. turns the lion into a 3d model	3858595085.mp4
Seed: 4273030543 Prompt: Make the skater in the image suffer a fall	4273030543.mp4
Seed: 1747446564 Prompt: The monkey takes a cool look and then puts on sunglasses	1747446564.mp4

---

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/DeepRatAI/LTX-FastVideo-ZeroGPU_Optimized.git
cd LTX-FastVideo-ZeroGPU_Optimized

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Basic Usage

Text-to-Video

python inference.py \
  --prompt "A beautiful sunset over the ocean" \
  --height 704 \
  --width 1216 \
  --num_frames 121 \
  --seed 42

Image-to-Video

python inference.py \
  --prompt "The person in the image starts walking" \
  --conditioning_media_paths path/to/image.jpg \
  --conditioning_strengths 0.8 \
  --conditioning_start_frames 0 \
  --height 704 \
  --width 1216 \
  --num_frames 121 \
  --seed 42

---

🎛️ Parameters Guide

Parameter	Description	Default	Range
--height	Output video height	704	256-720
--width	Output video width	1216	256-1280
--num_frames	Number of frames	121	1-257
--frame_rate	FPS of output	30	1-60
--seed	Random seed	171198	Any integer
--guidance_scale	Prompt adherence	3.0	1.0-20.0
--num_inference_steps	Quality steps	50	1-100

Conditioning Parameters

• --conditioning_media_paths: Path(s) to conditioning images/videos
• --conditioning_strengths: Strength of each condition (0.0-1.0)
• --conditioning_start_frames: Frame index where condition starts

---

📊 Technical Details

Model Architecture

• Base Model: LTX Video (Lightricks)
• Precision: Mixed (BF16/FP32)
• VAE: Causal Video Autoencoder
• Transformer: 3D Transformer with symmetric patchifier
• Scheduler: Rectified Flow

System Requirements

• GPU: NVIDIA GPU with 16GB+ VRAM (recommended)
• RAM: 32GB+ recommended
• Storage: 50GB+ for models
• Python: 3.10+

Supported Resolutions

• Width: 256px - 1280px (divisible by 32)
• Height: 256px - 720px (divisible by 32)
• Frames: 1 - 257 (formula: N * 8 + 1)

---

💡 Tips for Best Results

Text-to-Video

• Be specific about motion, lighting, and camera movement
• Use descriptive language: “slowly”, “dramatic”, “cinematic”
• Start with lower resolutions for faster iteration
• Avoid overly complex or contradictory prompts

Image-to-Video

• Use conditioning strength 0.7–0.9 for natural motion
• Clear, high-quality input images work best
• Describe the desired motion explicitly
• Avoid very low conditioning strength (<0.5)

General Tips

• Use negative prompts to avoid unwanted elements
• Adjust guidance scale: lower (2–4) for creativity, higher (5–8) for accuracy
• More inference steps = better quality but slower
• Use consistent seeds to reproduce results

---

📁 Project Structure

LTX-FastVideo-ZeroGPU_Optimized/
├── inference.py              # Main inference script
├── app.py                    # Gradio web interface
├── configs/                  # Configuration files
│   └── ltxv-13b-0.9.7-dev.yaml
├── ltx_video/               # Core LTX Video modules
│   ├── models/
│   ├── pipelines/
│   └── schedulers/
├── examples/                # Example outputs
│   ├── PtV/                # Picture-to-Video examples
│   └── ItV/                # Image-to-Video examples
└── requirements.txt         # Python dependencies

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

🙏 Credits & Acknowledgments

• LTX Video: Lightricks
• Model: Lightricks/LTX-Video
• Paper: LTX-Video: Realtime Video Latent Diffusion
• Interface: Built with Gradio
• Maintainer: DeepRat (solo)

---

🔗 Links

• 🌐 Hugging Face Space
• 📦 Model Card
• 📖 Original Repository
• 📄 Research Paper

---

🐛 Known Issues & Limitations

• Very high resolutions (>1280×720) require significant VRAM
• CPU inference is extremely slow (GPU strongly recommended)
• Long prompts (>77 tokens) may be truncated
• Some complex motions may not be fully captured

---

Made with ❤️ by the DeepRat for the Community

⭐ Star us on GitHub — it helps!

⬆ Back to Top