Upload a photo, pick a piece, see it on you — in seconds.
A full-stack virtual try-on tool for jewellery: pick an item from the catalogue, upload a photo, and get a photorealistic AI-generated image of yourself wearing it.
- Choose a piece from the catalogue — necklaces, rings, earrings, bracelets.
- Upload a photo — the app automatically asks for a face/upper-body photo or a hand photo depending on what you picked.
- See the reveal — the backend builds a detailed prompt from the item's material and placement, and generates a real image of you wearing it.
- Backend: FastAPI (Python)
- Image generation: Hugging Face Inference API — FLUX.1 Kontext, free tier, no billing required
- Frontend: React + Vite
- Storage: local filesystem only, no database
cd backend
pip install -r requirements.txtCreate backend/.env:
HF_TOKEN=your_huggingface_token_here
MOCK_MODE=false
Get a free token at huggingface.co/settings/tokens, and accept the license on the FLUX.1-Kontext-dev model page before your token can use it.
Don't want to grab a token just to look around? Set MOCK_MODE=true instead — the app runs the full flow end-to-end and returns a placeholder result, no API calls made.
uvicorn app:app --reloadRuns at http://localhost:8000.
cd frontend
npm install
npm run devRuns at http://localhost:5173.
prompts.py is the core of this project. The model is framed as a retoucher, not a generator — the prompt explicitly says this is a precise editing task, not a creative one, which keeps the model from redesigning the user's face or the jewellery instead of just compositing them. Each jewellery type also gets its own spatial placement instructions (how a ring wraps a finger, how a chain drapes a collarbone) and material-specific rendering hints (how gold catches light vs. how a pearl scatters it), because vague instructions produce vague results.
This started as a Gemini-powered pipeline — genuinely great results, until Google's free-tier image quota got cut to zero mid-build. Rather than shelve the project, the backend now runs on Hugging Face's free FLUX.1 Kontext model instead. The trade-off: FLUX edits based on the text description of a piece rather than seeing the exact product photo, so results are a little looser than Gemini's were — a fair price for zero cost.
| Feature | Status |
|---|---|
| Catalogue loading | working |
| Photo upload with preview | working |
| Correct photo type per jewellery type | working |
| Prompt construction | working |
| Image generation (Hugging Face FLUX Kontext) | working |
| Mock mode for API-free demos | working |