LTX Video

✨Key Features

• ▸Core Architecture & Capabilities: DiT-based model supporting text-to-video, image-to-video, video-to-video, multi-keyframe conditioning, keyframe animation, forward/backward video extension, and up to 60-second clips; LTX-2 adds native synchronized audio + video in one pass.
• ▸High Performance & Quality: Generates native 4K resolution videos at up to 50 FPS; distilled and FP8 quantized versions enable real-time inference (15× faster, runs on consumer GPUs/H100); supports resolutions divisible by 32 and frame counts like 17, 25, 257.
• ▸Multiple Model Variants: Includes high-quality 13B models (ltxv-13b-0.9.8-dev), faster distilled versions (13B & 2B), FP8 quantized for low VRAM, plus control models (depth, pose, canny) and LoRA fine-tuning support.
• ▸Open-Source & Integrations: Full inference code under Apache-2.0 license; model weights on Hugging Face under OpenRail-M (commercial use allowed); seamless with ComfyUI (custom nodes/workflows), Diffusers library, and community extensions (e.g., TeaCache for 2× speed).
• ▸Usage & Setup: Easy local installation via pip/venv; CLI inference scripts, Python library usage; online demos available on LTX Studio app, fal.ai, and Replicate for quick testing without setup.
• ▸Development Status: Active project by Lightricks (creators of Facetune & LTX Studio); repository now redirects focus to LTX-2 (announced 2025) for advanced audio-video features; weights/code for LTX-2 progressively released.
• ▸Target Audience: Developers, researchers, studios, and creators who want local, customizable, high-control video gen without cloud dependency; ideal for prototyping, fine-tuning, or integrating into pipelines.
• ▸Limitations: Best results under 720×1280 & <257 frames; requires specific input formats (divisible by 32/8+1); prompt adherence improves with detailed, chronological descriptions.