NVIDIA has unveiled 'LongLive-2.0,' a real-time video generation AI that achieves lightweight and high-quality generation through training designed for FP4 quantization.
NVIDIA's AI research team has unveiled ' LongLive-2.0, ' an AI model for generating videos. LongLive-2.0 is designed for real-time and consistent video generation over long periods of time, and achieves low memory usage and high accuracy by being designed with NVFP4 quantization in mind.
LongLive-2.0
LongLive 2.0 - YouTube
AI models can reduce memory usage by using a technique called 'quantization.' NVIDIA's Blackwell-generation and later GPUs can run AI models quantized to FP4, but existing AI models have faced the challenge of quality degradation when quantized to FP4. LongLive-2.0 is designed with NVFP4 quantization in mind from the training stage, reducing memory usage while minimizing quality degradation.
The images below show the video generation results using BF16 precision on the left and NVFP4 precision on the right. While the texture quality is rougher with NVFP4 precision, the prompt instructions were followed.
The NVFP4 quantized version, LongLive-2.0, uses only 19.4GB of memory and boasts a generation speed 1.84 times faster than the base model.
The difference in the generated results between the base model and the NVFP4 quantized version can be seen in the example images posted on
LongLive-2.0 is available as an open model in three forms: the 'Base Model,' the 'NVFP4 Quantized 4-Step Generative Model,' and the 'NVFP4 Quantized 2-Step Generative Model.' The license is the NVIDIA Open Model License .
Efficient-Large-Model/LongLive-2.0-5B · Hugging Face
https://huggingface.co/Efficient-Large-Model/LongLive-2.0-5B
Efficient-Large-Model/LongLive-2.0-5B-NVFP4-S4 · Hugging Face
https://huggingface.co/Efficient-Large-Model/LongLive-2.0-5B-NVFP4-S4
Efficient-Large-Model/LongLive-2.0-5B-NVFP4-S2 · Hugging Face
https://huggingface.co/Efficient-Large-Model/LongLive-2.0-5B-NVFP4-S2
Related Posts:
