kuzyn8/LTX-Video-0.9.5-diffusers

LTX-Video in Rust (Candle)

This repository provides a high-performance, native Rust implementation of LTX-Video using the Candle ML framework.

Features

• 🦀 Native Rust: No Python dependency required for inference.
• 🚀 Performance: Optimized for NVIDIA GPUs with Flash Attention v2 and cuDNN.
• 💾 Memory Efficient: Supports GGUF quantization for T5-XXL text encoder and VAE tiling/slicing for generating HD videos on consumer GPUs.
• 🛠 Flexible: Easy to use CLI for video generation and library for custom integration.

Quick Start

Installation

Ensure you have Rust and the CUDA Toolkit installed, then:

git clone https://github.com/FerrisMind/candle-video
cd candle-video
cargo build --release --features flash-attn,cudnn

Video Generation

cargo run --example ltx-video --release -- \
    --local-weights ./models/ltx-video \
    --prompt "A serene mountain lake at sunset, photorealistic, 4k" \
    --width 768 --height 512 --num-frames 97 \
    --steps 30

Performance & Memory

Resolution	Frames	VRAM (BF16)	VRAM (VAE Tiling)
512x768	97	~8-13 GB	~8-9 GB

Note: Using GGUF T5 encoder saves an additional ~8-12GB of VRAM.

Credits

• Original Model: Lightricks/LTX-Video
• Framework: HuggingFace Candle
• T5 v1_1 XXl GGUF and Safetensors: city96/LTX-Video-gguf (for GGUF support patterns, T5 XXl GGUF and Safetensors)

---

For more details, visit the main GitHub Repository.