A web application that transforms RAW camera files into professionally-edited photos using a three-stage ML pipeline: Denoise → Sharpen → Color Grade.

• Frontend: React 18, TypeScript, Tailwind CSS, Vite
• Backend: Python, FastAPI, OpenCV, rawpy
• ML Pipeline: DnCNN (denoise), Real-ESRGAN (sharpen), DPED CNN (color grade)
• Processing: Patch-based for large images, session caching for RAW decode

cd raw-enhance
pip install -r requirements.txt
python -m uvicorn backend.main:app --reload --host 0.0.0.0 --port 8000

cd frontend
npm install
npm run dev

Open http://localhost:5173 — the frontend proxies API calls to port 8000.

POST /enhance        Upload RAW + profile → enhanced JPEG
GET  /profiles       List available profiles
GET  /health         Health check

RAW File → rawpy decode → float32 RGB [0,1]
  → Stage 1: Denoise (bilateral filter / DnCNN)
  → Stage 2: Sharpen (unsharp mask / Real-ESRGAN)
  → Stage 3: Color Grade (parametric curves / DPED CNN)
  → JPEG output

The training script uses the DPED (DSLR Photo Enhancement Dataset) — freely downloadable, no registration required.

1. Download the dataset

Go to http://people.ee.ethz.ch/~ihnatova/ and download one of the phone device archives:

2. Train (run from inside raw-enhance/)

cd raw-enhance

# Train on iPhone pairs
python3 backend/training/train.py \
  --dataset ../dataset/iphone \
  --device iphone \
  --crop-size 100 \
  --epochs 50 \
  --batch-size 16

# Or train on BlackBerry / Sony pairs instead
python backend/training/train.py \
  --dataset ../dataset/blackberry \
  --device blackberry \
  --crop-size 100 \
  --epochs 50 \
  --batch-size 16

3. Evaluate a checkpoint

python backend/training/train.py \
  --dataset ../dataset/iphone \
  --device iphone \
  --crop-size 100 \
  --eval-only \
  --checkpoint weights/colorgrade.pth

The trained colorgrade.pth file goes in backend/weights/ and is loaded automatically when PIPELINE_MODE=pytorch.

The original plan used the MIT-Adobe FiveK dataset (5,000 RAW photos with professional retouches). It requires submitting an academic request and can take days to get approved. Switched to DPED (DSLR Photo Enhancement Dataset), which is freely downloadable with no sign-up at http://people.ee.ethz.ch/~ihnatova/. DPED has ~160,000 pre-cropped 100×100 phone/DSLR image pairs, making it better suited for patch-based CNN training anyway.

PyTorch's DataLoader with num_workers > 0 spawns child processes and pickles everything it sends to them. In Python 3.13, macOS switched from fork to spawn as the default multiprocessing start method. spawn requires all objects to be picklable by name — but collate_fn was defined as a local function inside train(), which can't be pickled. Fix: moved collate_fn to module level (_collate_fn) and set num_workers=0 so loading runs in the main process, bypassing spawn entirely.

With 100×100 input patches, three MaxPool2d(2) layers halve dimensions: 100 → 50 → 25 → 12 (odd number floors). The decoder then doubles back: 12 → 24 → 48 → 96 — 4 pixels short of the 100×100 target, causing a RuntimeError in the L1 loss. Fix: pad the input to the next multiple of 8 (= 2³ pool stages) at the start of forward, process at 104×104, then crop the output back to the original 100×100. This makes the UNet resolution-agnostic.

The backend tried to write temporary RAW files and session cache to the mounted workspace folder (/mnt/editing_ai/), which has restricted write permissions. Fix: switched all temp file paths to tempfile.gettempdir() (/tmp), which is always writable.

npm install on an arm64 Mac pulled in @rollup/rollup-linux-x64-gnu instead of the arm64 variant, causing a MODULE_NOT_FOUND error at build time. Fix: npm install @rollup/rollup-linux-arm64-gnu --save-dev.

• PIPELINE_MODE=opencv — OpenCV fallback (default, runs anywhere)
• PIPELINE_MODE=pytorch — Real ML models (requires GPU + model weights)

• Frontend: cd frontend && npm run build → deploy dist/ to Vercel
• Backend: Deploy FastAPI to Modal or Render with PIPELINE_MODE=pytorch