add TensorRT guidance + updated speed numbers

README.md

@@ -28,3 +28,30 @@ Models, ONNX/ncnn files, training configs and the architecture live at
 **NERVE links:** [models & configs](https://huggingface.co/Phips/NERVE) · [try it online (ZeroGPU)](https://huggingface.co/spaces/Phips/nerve) · [convert a checkpoint](https://huggingface.co/spaces/Phips/nerve-convert) · [local app](https://github.com/Phhofm/nerve-app) · [train your own (Colab)](https://github.com/Phhofm/nerve-train)
 
 Runs on ZeroGPU; large images are processed in tiles. Apache-2.0.
+
+## Running it fastest (TensorRT)
+
+This Space runs on ZeroGPU with PyTorch. For **maximum speed on your own
+NVIDIA GPU**, export a static-shape fp16 ONNX and enable the two TensorRT
+builder options we found fastest (`builderOptimizationLevel=5` + CUDA graphs):
+
+```python
+import onnxruntime as ort
+
+sess = ort.InferenceSession(
+    "4x_NERVE_OTF_gan_1x3x480x640_fp16.onnx",
+    providers=[("TensorrtExecutionProvider", {
+        "trt_fp16_enable": True,
+        "trt_builder_optimization_level": 5,
+        "trt_cuda_graph_enable": True,
+        "trt_engine_cache_enable": True,
+        "trt_engine_cache_path": "trt_cache",
+        "trt_timing_cache_enable": True,
+    }), "CUDAExecutionProvider"],
+)
+```
+
+Measured on an RTX 3060 at `1x3x480x640` fp16 (mean of 50 runs): NERVE 2x
+**33.24 → 31.80 ms**, NERVE 4x **38.43 → 33.82 ms**. Full numbers, the SPAN_S /
+compact comparison, and notes: see the
+[release page](https://huggingface.co/Phips/NERVE#update--2026-09-30-tensorrt-builder-options-how-to-run-nerve-fastest).