SDXL gets a TensorRT turbo boost from NVIDIA

Anyone running local image generation knows that while Stable Diffusion XL produces gorgeous visuals, it is significantly heavier on compute than earlier models.

Stability AI just announced a partnership with NVIDIA to give SDXL a massive speed injection. By integrating NVIDIA's TensorRT deep learning inference framework, the company has released optimized ONNX weights globally to make image generation radically faster.

On NVIDIA's flagship H100 chips, SDXL's performance has effectively doubled, churning out full high-definition images in just 1.47 seconds.

The bigger the GPU, the bigger the gains

Basically, TensorRT strips out inefficiency by fusing layers and tailoring deep learning execution directly to NVIDIA hardware architectures.

Stability AI shared detailed benchmarks across three common enterprise accelerators: A10, A100, and H100.

In terms of latency reduction over the baseline unoptimized model, TensorRT sped up the A10 by 13%, the A100 by 26%, and the H100 by 41%.

Throughput told an even more dramatic story. The optimized model pushed 20% higher throughput on the A10, 33% more on the A100, and a massive 70% increase on the H100.

Stability AI also noted that this is just step one, with 8-bit precision optimizations already in the pipeline to cut resource demands even further.

Why it matters

If you run an AI image generation product or process high volumes of graphics, this is an instant bottom-line win.

Getting latency down under 1.5 seconds while boosting throughput by 70% means you can either cut your GPU cloud rental bills nearly in half or serve almost double the concurrent users on the same infrastructure.

In the broader generative AI race, this keeps open-source weights competitive against closed-source titans like Midjourney. When the open ecosystem can deliver photorealism at production speeds, building on open models becomes a much easier decision for startups.

We think leaning directly on NVIDIA's low-level toolchain is the smartest move Stability AI could make here. Instead of trying to reinvent execution logic, they let the hardware king optimize the pipe.

The only real catch is that TensorRT naturally tightens NVIDIA's walled garden across cloud AI infra, leaving alternative hardware options even further in the dust.