Visual Generation (Beta)

This section documents optimization features for VisualGen image and video generation. Use the guides below for configuration, supported options, and interactions between features.

For supported models and an overview of VisualGen, see the model overview. For usage examples, see VisualGen Examples.

• CUDA Graphs

• Quantized Attention

• Sparse Attention

Related technical blogs

These articles explain the design choices and benchmark results behind VisualGen optimizations:

• Scaling video generation across NVL72 covers multi-GPU parallelism and scaling results.

• Quantization and skipsoftmax attention for video generation covers GEMM quantization, quantized attention, and Skip Softmax Attention, including performance and visual-quality trade-offs.