Stable Diffusion, but you pre-bake the neural network first

What it does The extension acts as a bridge between NVIDIA’s TensorRT SDK and the Automatic1111 Stable Diffusion Web UI. It generates optimized inference engines tailored to specific checkpoints, resolutions, and batch sizes, then hooks them into the standard generation pipeline. Supported model families include SD 1.5, 2.1, SDXL, SDXL Turbo, and LCM.

The interesting bit The rigidity is the feature. TensorRT engines are hyperspecialized: static ones lock you to a single resolution and batch size, while dynamic ones cover a range at the cost of extra VRAM and a small performance hit. The README even warns that using hires.fix requires building an engine that spans both the starting and ending resolutions, because the compiled graph cannot simply wing it.

Key highlights

• Supports SD 1.5, 2.1, SDXL, SDXL Turbo, and LCM checkpoints, though SDXL variants prefer 12 GB+ of VRAM.
• Generates static engines for exact resolutions and batch sizes, or dynamic engines that trade a little performance for flexibility.
• LoRA and LyCORIS checkpoints require a separate TensorRT weight conversion step that takes roughly 20 seconds.
• First-time engine generation for a checkpoint is slow; subsequent engines for the same checkpoint build much faster.
• All output resolutions must be multiples of 64, including both low- and high-res stages when using hires.fix.

Caveats

• Several popular Web UI flags—medvram, lowvram, and api—are known to break engine compilation or model registration.
• Using hires.fix demands a dynamic engine that explicitly covers both the initial and upscaled resolutions.
• If the TensorRT tab fails to appear after installation, the README attributes this to a failed install and points to a manual workaround.

Verdict Worth a look if you run Automatic1111 on an NVIDIA RTX GPU and want to trade setup time for inference speed. Skip it if you depend on medvram or lowvram flags, or regularly generate at resolutions that are not multiples of 64.