Stability AI and NVIDIA Ship Optimized Stable Diffusion 3.5 NIM for Enterprises

Stability AI has launched a new NVIDIA NIM microservice for Stable Diffusion 3.5 Large, promising significantly faster image generation and simplified enterprise deployment.

The SD3.5 NIM packages the model, ControlNets, and optimized inference engines into a single container. In NVIDIA's H100 GPU benchmarks, the TensorRT-optimized service generated images in 3,700 milliseconds — 1.8x faster than the 6,800ms baseline using standard PyTorch. The microservice supports Depth and Canny ControlNets within the same container, eliminating the need for separate deployments for each variant.

For enterprise creative teams, the speed gains translate to faster iteration cycles, larger batch processing, and the ability to run multiple models simultaneously — workflows that previously hit practical limits on existing hardware. The service runs on NVIDIA's Ada and Blackwell data center GPUs.

The microservice is available through NVIDIA's build.nvidia.com platform, with model weights downloadable from Hugging Face. Smaller organizations and researchers can use the optimized models under Stability AI's permissive community license. Enterprises with annual revenue over $1 million need to contact Stability AI for a commercial license, which includes implementation support and customization options.