# CNN v1: Original Post-Processing Neural Network

**Architecture:** 3-layer convolution, generated shader weights
**Status:** Active (used in timeline), legacy architecture

## Overview

Original CNN implementation with per-layer WGSL shaders. Supports multiple kernel sizes (1×1, 3×3, 5×5, 7×7) with generated weight arrays.

**For new work, use CNN v2** (`cnn_v2/`) which provides:
- Storage buffer architecture (~3.2 KB vs generated WGSL)
- 7D static features (RGBD + UV + sin + bias)
- Sigmoid activation with stable training
- Dynamic layer configuration

## Quick Reference

**Training:**
```bash
./cnn_v1/training/train_cnn.py --input training/input --target training/output \
  --layers 3 --kernel_sizes 3,5,3 --epochs 5000
```

**Integration:**
- **C++:** `cnn_v1/src/cnn_effect.{h,cc}`
- **Assets:** `workspaces/main/assets.txt` (lines 40-46)
- **Timeline:** `workspaces/main/timeline.seq` (CNNEffect)

## Documentation

- [CNN.md](docs/CNN.md) - Architecture overview
- [CNN_V1_EFFECT.md](docs/CNN_V1_EFFECT.md) - Implementation details
- [CNN_TEST_TOOL.md](docs/CNN_TEST_TOOL.md) - Testing guide
- [CNN_DEBUG.md](docs/CNN_DEBUG.md) - Debugging notes

## Directory Structure

```
cnn_v1/
├── README.md              # This file
├── src/
│   ├── cnn_effect.h       # Effect header
│   └── cnn_effect.cc      # Effect implementation
├── shaders/               # WGSL shaders (7 files)
├── training/              # Python training script
└── docs/                  # Documentation (7 markdown files)
```

## Differences from CNN v2

| Feature | CNN v1 | CNN v2 |
|---------|--------|--------|
| Architecture | Generated WGSL weights | Storage buffer weights |
| Input Features | 4D (RGBA/prev layer) | 12D (4D + 8D static) |
| Activation | ReLU | Sigmoid + ReLU |
| Size | ~Variable (WGSL gen) | ~3.2 KB (binary) |
| Training | Full-image | Patch-based (default) |
| Layer Config | Compile-time | Runtime (dynamic) |

## Migration Notes

CNN v1 remains in the timeline for historical validation. For new effects or experiments, use CNN v2's enhanced feature set and compact binary format.

See `cnn_v2/docs/CNN_V2.md` for CNN v2 architecture details.