Enhanced Stable Diffusion 3.5 Models with TensorRT: Achieving 2X Speed and 40% Memory Savings on NVIDIA RTX GPUs

SD3.5 has made significant strides in performance and accessibility, establishing itself as a leading customizable image model. It excels in generating diverse styles, from photography to 3D art, and produces images reflecting a variety of skin tones and features with minimal prompting. SD3.5 Large stands out for its exceptional prompt adherence, accurately following text inputs for high-quality output. The integration of TensorRT optimization enhances efficiency across NVIDIA RTX GPUs, allowing for a reduction in model size via FP8 quantization, which preserves output quality. This optimization enables SD3.5 Large to run on five RTX 50 Series systems concurrently, a significant increase from one system previously. Enhanced performance is evident, with generation speeds improving by up to 2.3x for SD3.5 Large and 1.7x for SD3.5 Medium, while also reducing VRAM requirements by 40%.

Source link