ltx-video

LTX-Video is an open-source DiT-based video generation model developed by Lightricks. It integrates core video generation capabilities into a single model, including synchronized audio and video generation, high-fidelity outputs, and multiple performance modes. The model can generate videos at native 4K resolution up to 50 FPS with synchronized audio in one pass. Key capabilities include: - Text-to-video and image-to-video generation - Video extension (forward and backward) - Multi-keyframe conditioning and keyframe-based animation - Video-to-video transformations - Synchronized audio generation (in LTX-2) The repository offers multiple model variants, including 13B and 2B parameter versions, distilled models for faster inference, and FP8 quantized versions for reduced VRAM usage. Distilled models support rapid iteration, generating HD videos in seconds on appropriate hardware. The 2B distilled model requires only 1GB of VRAM in its LoRA variant. LTX-Video integrates with ComfyUI, Diffusers, and provides a Python library API. It supports control models for depth, pose, and Canny edge conditioning, as well as LoRA fine-tuning for style customization. The model is available under the OpenRAIL-M license for commercial use. Community contributions include ComfyUI-LTXTricks for advanced editing workflows (RF-Inversion, RF-Edit, FlowEdit), LTX-VideoQ8 for 8-bit optimized inference on NVIDIA ADA GPUs, and TeaCache for training-free inference acceleration. The successor model, LTX-2, adds synchronized audio-video generation, longer clip support, and enhanced creative controls.