Stability AI partners with NVIDIA to deliver optimized versions of Stable Diffusion 3.5

Stability AI, working in collaboration with NVIDIA, has introduced optimizations for the Stable Diffusion 3.5 model family using both TensorRT and floating-point 8 (FP8) precision. These upgrades improve generative AI performance on RTX GPU hardware by significantly enhancing image generation speed while reducing video RAM (VRAM) requirements. The enhancements build on Stable Diffusion 3.5’s original design for consumer hardware, extending its reach for creative professionals and developers regardless of their workstation specs.

With these TensorRT-based improvements, users of Stable Diffusion 3.5 Large can expect generation speeds up to 2.3 times faster, while the Medium version sees a 1.7 times increase. VRAM needs are reduced by up to 40 percent, allowing more complex or higher-resolution workloads on compatible hardware. While performance upgrades benefit technical users, broader accessibility across diverse setups is a key outcome, particularly for those working on limited or consumer-grade GPUs.

The TensorRT-optimized model weights are distributed under the Stability AI Community License, making them available for both commercial and non-commercial use. Users can access the models through Hugging Face, with associated deployment code available from NVIDIA’s GitHub repository.