Spaces:
Sleeping
Sleeping
File size: 2,749 Bytes
3a330e0 6025e8c 3b8e240 3a330e0 6025e8c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | ---
title: Luxury Authenticator
emoji: 💎
colorFrom: gray
colorTo: yellow
sdk: gradio
sdk_version: 6.14.0
python_version: 3.11
app_file: app.py
pinned: false
---
# Luxury Authenticator
**One image. Any source. Structured signals.**
A 5-layer AI pipeline (research demo) that analyses a luxury-related image and returns a structured report: image source type, item identification, a CLIP-based confidence signal, a perceptual-hash provenance check against a small demo database, and rule-based recommended actions.
## Architecture
| Layer | Purpose | Model / method |
|-------|---------|----------------|
| 1 — Image type | Real photo / AI / Screenshot / Render | CLIP zero-shot (`openai/clip-vit-base-patch32`) |
| 2 — Object ID | Brand, category, caption | BLIP-base + CLIP (`Salesforce/blip-image-captioning-base`) |
| 3 — Confidence signal | Visual similarity vs authentic vs replica wording | CLIP zero-shot (same CLIP) |
| 4 — Provenance | Near-duplicate check vs demo scam list | `imagehash` pHash + `data/scam_database.csv` |
| 5 — Actions | Short recommendations | Rule engine (no LLM) |
## Honest limitations
- The confidence signal (Layer 3) uses CLIP visual similarity, not professional authentication.
- The provenance database contains a small set of manually seeded demo entries (hashes from `data/flagged/`).
- Do not use for purchase decisions over INR 50,000 without professional verification.
## Local run
```bash
pip install -r requirements.txt
python app.py
```
First request downloads PyTorch and model weights (expect several minutes and ~2 GB RAM peak on CPU).
## Rebuild scam CSV
After adding PNG/JPG files under `data/flagged/`:
```bash
python data/seed_phashes.py
```
## Demo assets
Example images live in `examples/`. Flagged images used only for hashing live in `data/flagged/`. Uploading `examples/fb_listing_screenshot.png` should match the duplicate entry in the demo database and show **Flagged** in Layer 4.
## Timing (reference)
End-to-end inference target is under ~25 seconds per image on Hugging Face Spaces CPU after models are loaded; cold start adds model download and load time on the first request.
## Smoke tests
- Layer 4 only (no PyTorch): `set PYTHONPATH=%CD%` then `python scripts/smoke_layer4.py` (Windows) or `PYTHONPATH=. python scripts/smoke_layer4.py` (Unix).
- Full pipeline (requires `pip install -r requirements.txt`): `python scripts/smoke_analyse.py` from the repo root (prints timing and a short summary for `examples/chanel_bag.png`).
## V2 ideas
GradCAM / heatmaps, TinEye-style API for provenance, larger brand lists, calibrated thresholds, optional LLM action layer.
## Built by
IT portfolio project — Luxury Truth Lens PRD v2.0 (Buildable Edition).
|