Sure! Here’s the translation:
Generative artificial intelligence is experiencing a true revolution, and once again, NVIDIA is at the forefront with a move that promises to change the game: the optimization of the Stable Diffusion 3.5 Large model alongside Stability AI, one of the world’s most popular AI image engines. The result: a significantly more efficient model that consumes 40% less memory and doubles performance on the latest RTX graphics cards, making high-quality image generation accessible to more users and creators than ever before.
Why does this improvement matter? The VRAM bottleneck
Until now, working with advanced image generation models like Stable Diffusion 3.5 Large was a privilege reserved for those with high-end equipment, due to the high video memory (VRAM) requirements. The original version required more than 18 GB of VRAM, making it inaccessible for most freelancers, digital artists, small studios, and hobbyists. Only data centers or specialized workstations could harness its full potential.
Thanks to FP8 quantization, an advanced technique that allows models to run with lower precision in calculations without sacrificing visual quality, NVIDIA and Stability AI have reduced this requirement to 11 GB of VRAM. This means that much more affordable graphics cards—like the new GeForce RTX 50 Series—can run these models locally, paving the way for a true democratization of generative AI.
TensorRT: The “turbo” for visual artificial intelligence
The key to this advancement is TensorRT, NVIDIA’s inference acceleration software, which optimizes the performance of AI models to make the most of the dedicated cores (Tensor Cores) in RTX graphics cards. This “turbo” has not only reduced memory usage but has also multiplied the speed of image generation.
- SD3.5 Large FP8 with TensorRT:
+2.3 times faster than the traditional version in PyTorch BF16
-40% memory usage - SD3.5 Medium:
+1.7 times acceleration compared to PyTorch
Users can now generate high-resolution images with the same quality, but in half the time and with significantly lower hardware demands. Most importantly, this advancement can be leveraged on more than 100 million RTX PCs and workstations
A growing ecosystem: Hugging Face, NIM microservices, and more
The models optimized by TensorRT and quantized to FP8 are already available on Hugging Face, one of the most popular open AI platforms, allowing direct download and usage for both developers and advanced users.
But the commitment goes much further. NVIDIA and Stability AI are preparing to launch Stable Diffusion 3.5 as a NIM microservice, which will further facilitate its deployment on servers, applications, and cloud platforms, enabling easy integration into professional workflows and business solutions. It is expected to be available in July, which will enhance the arrival of generative AI in new scenarios: from editing apps, video games, and film to industrial, scientific, and medical applications.
TensorRT for RTX: Standalone SDK and easier implementation
During the Microsoft Build conference, NVIDIA presented the new standalone TensorRT SDK for RTX, which removes many barriers for developers. Previously, it was necessary to pre-generate optimized engines for each class of GPU, which slowed down deployment. Now, the system generates the ideal engine for each GPU “in real time” (JIT compilation), simplifying integration into Windows applications thanks to direct support with Windows ML.
This reduces the SDK size, facilitates portability across devices, and ensures that any user can benefit from the latest AI optimizations with a simple update.
Generative AI for everyone: NVIDIA’s vision and the future of the industry
This improvement arrives in a time of true boom for generative AI: from the creation of photorealistic images, concept art, industrial design, or content for video games, to the automatic generation of videos, music, and text. Companies like Adobe, Autodesk, Canva, and dozens of startups have already begun integrating generative models into their tools, fundamentally transforming creativity and digital productivity.
With the arrival of optimized Stable Diffusion 3.5, access to cutting-edge generative AI is more global, efficient, and affordable. The collaborative development between NVIDIA and Stability AI, along with partnerships with platforms like Hugging Face and the push for portability through NIM microservices, paints a picture where any professional, studio, or SME can deploy advanced visual AI without relying on the cloud or extensive infrastructure.
Additionally, NVIDIA wanted to showcase its strength in the professional sector with these announcements at GTC Paris and VivaTech, the leading European forums for innovation, where Jensen Huang, the company’s CEO, emphasized the importance of generative AI as a driver of competitiveness and transformation across all industries, from creative sectors to healthcare, mobility, and education.
source: Artificial intelligence news