Stability AIとNVIDIA、Stable Diffusion 3.5 NIMで高速化・企業向け展開を簡素化

Key Takeaways:

Stability AI is launching a new NVIDIA NIM microservice for Stable Diffusion 3.5, making it faster and simpler for enterprises to deploy our most advanced models.

The SD3.5 NIM is available at build.nvidia.com. Users can download model weights directly from Hugging Face.

The models are available for commercial and non-commercial use under our permissive Stability AI Community License. For enterprises with annual revenue over $1M, please contact us to discuss our Enterprise Licensing, which offers additional support and customization options.

image

We're excited to announce our collaboration with NVIDIA to launch the Stable Diffusion 3.5 NIM microservice, enabling significant performance improvements and streamlined enterprise deployment for our leading image generation models.The SD3.5 NIM supports Stable Diffusion 3.5 Large, with expanded model compatibility planned for future releases.

The SD3.5 NIM delivers faster image generation on enterprise hardware. This enables complex image generation workflows that weren’t as feasible before.

A faster, easier way for enterprises to deploy Stable Diffusion models

A NIM provides a simplified, optimized way to run AI inference by packaging inference engines, APIs, and model configurations into secure, portable containers. Think of it as a pre-configured, enterprise-ready package that eliminates the complexity of setting up and optimizing AI models from scratch.

The SD3.5 NIM delivers performance gains that improve efficiency and ease of deployment for enterprises:

Speed improvements: 1.8x performance gains over PyTorch, with NVIDIA H100 GPUs testing showing TensorRT-optimized generation at 3,700ms compared to 6,800ms for standard PyTorch on SD3.5 Large.

Consolidated deployment: The SD3.5 NIM supports the SD3.5 Large model with Depth and Canny ControlNets within a single container, meaning that instead of needing separate deployments for each model, users get all versions packaged together.

The SD3.5 NIM supports enterprise and data center Ada and Blackwell GPUs.

Advanced workflows made practical for enterprise creative teams

These optimizations enable faster iteration cycles, larger batch processing, and more complex workflows that were previously less practical due to hardware limitations.

The efficiency gains are particularly valuable for advanced workflows, such as running multiple models simultaneously. The performance improvements make complex, multi-model workflows more feasible, opening new possibilities for advanced users and making rapid scaling a simpler option when utilizing cloud deployments.

Get started

The SD3.5 NIM is available at build.nvidia.com. Users can download Stable Diffusion 3.5 model weights directly from Hugging Face.

For smaller organizations and researchers getting started, the optimized models are available for commercial and non-commercial use under the permissive Stability AI Community License.

For enterprises with annual revenue over $1M, please contact us to discuss our Enterprise Licensing, with implementation support, customization options and professional services available. You can also visit Stability AI Solutions to learn more about customizing models and workflows for specific use cases.

To stay updated on our progress, follow us on X, LinkedIn, Instagram, and join our Discord Community.