README: replace ASCII architecture diagram with docs/scheme.jpg

Graphic beats monospace art for the Architecture section: more
legible on mobile, renders fine on GitHub + any markdown viewer,
same information (input -> calibration pipeline -> runtime engine
-> output). Kept a short 4-bullet prose caption under the image
so the content is still searchable/accessible.

Author: bauratynov

README.md

@@ -151,45 +151,22 @@ ff.close()
 
 ## Architecture
 
-```
-┌───────────────────────┐       ┌───────────────────────────┐
-│   RGB face 112x112    │       │     Calibration pipeline  │
-│   fp32 [-1, 1] HWC    │       │   (one-shot, Python)      │
-└──────────┬────────────┘       │                           │
-           │                    │  ONNX model               │
-           │                    │      ↓                    │
-           │                    │  Per-channel activation   │
-           │                    │  statistics on LFW calib  │
-           │                    │      ↓                    │
-           │                    │  p99.9 percentile +       │
-           │                    │  outlier inclusion        │
-           │                    │      ↓                    │
-           │                    │  BN-fold + scale-fold     │
-           │                    │      ↓                    │
-           │                    │  w600k_r50_ffw4.bin (43M) │
-           │                    │  op_scales_v2.bin (0.2M)  │
-           │                    └───────────┬───────────────┘
-           │                                │
-           ▼                                ▼
-┌─────────────────────────────────────────────────────────────┐
-│                   fastface_int8.exe (96 KB)                 │
-│                                                             │
-│   ┌──────────────┐   ┌──────────────┐   ┌──────────────┐    │
-│   │ int8 Conv    │   │ int8 Conv    │   │ int8 Gemm    │    │
-│   │ + BN + PReLU │→→│ + Add (res)  │→→│ VNNI matvec  │    │
-│   │ fused        │   │ fused        │   │ 25088 → 512  │    │
-│   └──────────────┘   └──────────────┘   └──────────────┘    │
-│          ↑                                                  │
-│     Winograd F(2,3) Conv + AVX-VNNI vpdpbusd matvec         │
-│     Per-channel INT8 activation quant + weight fold          │
-└─────────────────────────────────────────────────────────────┘
-           │
-           ▼
-┌───────────────────────┐
-│   Embedding (512,)    │
-│   fp32, L2-able       │
-└───────────────────────┘
-```
+![FastFace architecture: input, calibration pipeline, runtime engine, output](docs/scheme.jpg)
+
+Three boxes:
+
+1. **Input** -- 112x112 RGB face, fp32 in [-1, 1], HWC layout.
+2. **Calibration pipeline** (offline, one-shot Python): ONNX model ->
+   per-channel activation statistics on LFW calib faces -> p99.9
+   percentile + outlier inclusion -> BN + scale fold -> produces
+   `w600k_r50_ffw4.bin` (43 MB) and `op_scales_v2.bin` (0.2 MB).
+3. **Runtime engine** `fastface_int8.exe` (96 KB): three fused
+   stages -- int8 Conv + BN + PReLU, int8 Conv + residual Add, int8
+   Gemm via AVX-VNNI `vpdpbusd` matmul (25088 -> 512). Winograd
+   F(2,3) for Conv 3x3, per-channel INT8 activation quant with
+   weight-fold.
+4. **Output** -- 512-d fp32 embedding, ready to L2-normalize for
+   cosine similarity.
 
 ---