NVIDIA Unveils 4K AI Video Generation on RTX GPUs

NVIDIA has announced a groundbreaking advancement in AI-powered video generation, enabling creators to produce high-quality 4K videos directly on PCs equipped with RTX GPUs. This breakthrough combines optimizations in PyTorch-CUDA, native NVFP4/FP8 precision support, and the newly released LTX-2 model from Lightricks, significantly enhancing performance and reducing VRAM usage.

The new pipeline, which includes updates to tools like ComfyUI and Llama.cpp, allows artists to generate videos up to 3x faster while maintaining sharp, clear visuals. This is achieved through the RTX Video node in ComfyUI, which upscales videos to 4K in real-time, sharpening edges and eliminating compression artifacts.

Accelerating AI Video Creation on PC

NVIDIA's collaboration with the open-source community has led to major performance improvements for small language models (SLMs) on RTX GPUs and the NVIDIA DGX Spark desktop supercomputer. These updates are particularly beneficial for mixture-of-experts models, including the new NVIDIA Nemotron 3 family of open models. The optimizations will be available in the next update of LM Studio and integrated into applications like the MSI AI Robot app, which leverages Llama.cpp enhancements to control MSI device settings.

The LTX-2 model, now available for download, stands out for its ability to generate up to 20 seconds of 4K video with impressive visual fidelity. It features built-in audio support, multi-keyframe capabilities, and advanced conditioning, providing creators with cinematic-level control without relying on cloud-based solutions.

Empowering Creators with Local AI Tools

The video generation workflow, set to release next month, includes the open weights of the LTX-2 model and ComfyUI RTX updates. This milestone for local AI video creation allows artists to turn storyboards into photorealistic keyframes and then into high-quality 4K videos. The pipeline is divided into three customizable blueprints, including Spark, which enables parallel asset generation while keeping the main PC available for editing.

NVIDIA has also improved ComfyUI's memory offload feature, known as weight streaming. This optimization allows the use of system RAM when VRAM is exhausted, enabling larger models and more complex node graphs on mid-range RTX GPUs. Additionally, the Hyperlink private beta, which offers a 35% faster inference performance for SLMs via Ollama and llama.cpp, is rolling out access starting this month.

These advancements, unveiled at CES 2026, mark a significant leap forward in generative AI on PC. With tools like the NVIDIA Broadcast app enhancing microphone and webcam quality for livestreaming and video conferencing, and the integration of RTX Video Super Resolution in ComfyUI, NVIDIA is paving the way for widespread adoption of AI video creation among creators, gamers, and productivity users.