Fine-tune SDXL with LoRA to generate paintings in Edward Hopper's style — dramatic lighting, isolation, geometric composition, muted palettes.
cd frontend
bun install --frozen-lockfile
bun run devuv sync --project scripts
uv run --project scripts python scripts/collect_images.py # Download Hopper paintings
uv run --project scripts python scripts/prepare_data.py # Resize + caption
uv run --project scripts modal run scripts/train.py # LoRA training on Modal GPU
uv run --project scripts modal run scripts/generate.py # Generate with trained LoRA- Base model: stabilityai/stable-diffusion-xl-base-1.0
- Training: LoRA via HuggingFace diffusers + peft
- Compute: Modal (A10G GPU)
- Budget: ~$1-2 total
- Frontend: Next.js app in
frontend/