add 'Why NERVE?' as the second section

README.md

@@ -39,6 +39,37 @@ Part of the **BODY** suite by **Philip Hofmann** — networks built to be
 
 ---
 
+## Why NERVE?
+
+There are many SR networks already. Here's the honest case for NERVE — and when
+to pick something else.
+
+**One-line pitch:** a ~1.8M-param, pure-conv SR net that trains with plain
+`train.py`, exports to dynamic ONNX on the first try, and converts to ncnn —
+with no fusion step, no variants, no custom ops.
+
+**Why it's worth your time:**
+- **One file, one variant.** The architecture is a single readable file and the
+  model is one size. No S/M/L matrix, no fused-vs-unfused checkpoints, nothing
+  to fuse before deploy — **what you train is what you ship.**
+- **Export just works.** Dynamic-shape ONNX (verified, ORT parity ~1e-5) and a
+  clean ncnn conversion. Standard ops only (`Conv / Add / ReLU / DepthToSpace /
+  Resize`).
+- **Boring, stable training.** Norm-free, bf16 AMP, no gradient clipping, a
+  single Charbonnier loss for the base stage.
+- **Small and fast.** ~1.8M params — SPAN-class size but conv-only, so it fits
+  mobile, edge, web and game engines.
+- **Real-world first.** Trained on a CC0 dataset with the full Real-ESRGAN OTF
+  degradation pipeline; the OTF GAN model restores degraded images, not just
+  clean bicubic downsamples.
+- **Maintainable for years.** One op set, no exotic dependencies, ~90-line arch.
+
+**When to pick something else:** if you want maximum PSNR/LPIPS on academic
+benchmarks, or a big transformer, NERVE isn't trying to win that — use HEART.
+NERVE wins on *simplicity, deployability and hassle-free training*.
+
+---
+
 ## Quickstart
 
 ### As a user (just want to upscale / restore images)