move + expand 'Why HEART?' to second section

README.md

@@ -36,6 +36,34 @@ real-world usable first, not benchmark-chasing. See
 
 ---
 
+## Why HEART?
+
+There are strong SR networks already. Here's the honest case for HEART — and
+when to pick something else.
+
+**One-line pitch:** HAT-class quality at roughly half the memory, in a single
+readable file that trains stably in bf16 and exports to dynamic ONNX.
+
+**Why it's worth your time:**
+- **Quality at lower cost.** Ties HAT on PSNR while using **~half the peak
+  VRAM**, and runs **~1.6x faster in bf16** inference.
+- **FlashAttention / SDPA friendly.** No attention masks, no position-bias
+  table lookups — RIB replaces them with a dot product.
+- **Stable, simple training.** i-LN keeps bf16 AMP stable with no fp32
+  gymnastics.
+- **One file, exports cleanly.** Single-file pure PyTorch; dynamic ONNX on the
+  first try; no fusion step or fused/unfused checkpoint pairs.
+- **Real-world capable.** OTF models are trained on the full Real-ESRGAN
+  degradation chain and restore real photos, not just clean bicubic downsamples.
+
+**When to pick something else:** for maximum benchmark PSNR or a tiny mobile
+model, use NERVE instead — HEART is the desktop / server quality tier.
+
+Full architecture details, design rationale, and ablation results:
+[`docs/EXPERIMENTS.md`](docs/EXPERIMENTS.md).
+
+---
+
 ## Quickstart
 
 ### As a trainer (want to train your own model)
@@ -107,22 +135,6 @@ All ONNX files are **dynamic-shape fp32, opset 17, onnxslim-optimized**.
 
 ---
 
-## Why HEART?
-
-- **~16.7M params**, ties HAT on PSNR at **~1.6× less VRAM** and faster.
-- **FlashAttention/SDPA friendly**: no attention masks, no position-bias table
-  lookups (RIB replaces them with a dot product).
-- **Stable bf16 AMP training** (i-LN) — no fp32 gymnastics.
-- **Single-file pure PyTorch**, exports cleanly to dynamic ONNX.
-- **Real-world capable**: OTF models trained on the full Real-ESRGAN
-  degradation chain restore lost texture on real photos, not just clean
-  bicubic downsamples.
-
-Full architecture details, design rationale, and ablation results:
-[`docs/EXPERIMENTS.md`](docs/EXPERIMENTS.md).
-
----
-
 ## Repository layout
 
 ```