OPP — Open Provenance Protocol

Shazam for AI-generated images.
An open protocol for minting and verifying provenance of AI-generated content.

Quick Start • How It Works • API • Protocol • Contributing

---

The Problem

AI generators produce billions of images. We need to know which images are AI-generated and by whom. Current solutions fall short:

Approach	Problem
C2PA / Content Credentials	Metadata embedded in file — stripped on screenshot, re-upload, compression
AI Detectors	Probabilistic guessing — unreliable, adversarial arms race
Blockchain provenance	Over-engineered — gas fees make billion-scale minting impractical
Proprietary watermarks	No interoperability — each generator's watermark is isolated

The Solution

OPP is an external, federated registry of AI image signatures. Think DNS for AI provenance.

• Generators mint signatures when creating images
• Anyone can verify by querying the registry
• Works after screenshots, crops, compression, filters — no embedded metadata to strip
• Open protocol — any generator, any verifier, any registry node

How It Works

3-Layer Signature System

Every image gets a signature with three complementary layers:

┌─────────────────────────────────────────────────────────┐
│  L1: SHA-256          → Exact duplicate detection       │
│  256-bit hash           O(1) instant lookup             │
├─────────────────────────────────────────────────────────┤
│  L2: PDQ (Meta)       → Near-duplicate matching         │
│  256-bit perceptual     Billion-scale, 2^256 values     │
│  hash                   Survives resize, compress, crop │
├─────────────────────────────────────────────────────────┤
│  L3: CLIP ViT-L/14    → Semantic similarity             │
│  768-dim embedding      Robust to all transformations   │
│                         Sub-10ms search at billions     │
└─────────────────────────────────────────────────────────┘

Matching Pipeline

Verification uses a tiered pipeline — fast to accurate:

1. SHA-256 exact match → O(1), instant
2. CLIP cosine similarity → Qdrant HNSW vector search, sub-10ms at billions
3. PDQ hamming distance → additional confidence signal on candidates

Adaptive Variant Tracking

OPP's registry gets smarter with every verification. When a query matches with high confidence, the variant's signature is automatically minted and linked to the original:

Original (minted by generator)
  ├── Screenshot (auto-minted on verify, 94% match)
  │     └── Crop of screenshot (auto-minted, 93% match)
  │           └── Re-upload of crop (auto-minted, 95% match)
  └── Compressed version (auto-minted, 96% match)

Each variant only needs to match its nearest ancestor, not the original. So even a heavily degraded image that's been screenshotted, cropped, and re-uploaded multiple times is caught — because the chain never breaks.

The most viral images (highest misuse risk) accumulate the most variants, making them the hardest to escape detection. The problem and the solution scale together.

Architecture

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  AI Generator   │     │  AI Generator   │     │  AI Generator   │
│  (OpenAI)       │     │  (Midjourney)   │     │  (Self-hosted)  │
│  Runs OPP lib   │     │  Runs OPP lib   │     │  Runs OPP lib   │
│  on own compute │     │  on own compute │     │  on own compute │
└────────┬────────┘     └────────┬────────┘     └────────┬────────┘
         │ push signature        │ push signature        │ push signature
         ▼                       ▼                       ▼
┌────────────────────────────────────────────────────────────────────┐
│                      OPP Central Index                            │
│                 (Qdrant + metadata store)                         │
│            Unified search across ALL generators                   │
└───────────────────────────┬────────────────────────────────────────┘
                            ▲
                            │ verify query (image upload)
                     ┌──────┴──────┐
                     │  Any Client │
                     │  (verifier) │
                     └─────────────┘

Quick Start

Option 1: Docker (Recommended)

git clone https://github.com/opp-protocol/opp.git
cd opp
docker compose up

This starts the OPP server + Qdrant vector database. Server runs on http://localhost:8000.

Option 2: Local Install

pip install opp-ai-protocol

Mint an Image

Generators compute signatures on their own infrastructure and push to the central index:

# Via CLI (computes signature locally, pushes to server)
opp mint my_image.png --generator "dall-e-3" --server http://localhost:8000

# Via API (pre-computed signature)
curl -X POST http://localhost:8000/v1/mint \
  -H "Content-Type: application/json" \
  -d '{
    "signature": {
      "protocol_version": "OPP/1.0",
      "l1_exact": { "algorithm": "sha256", "version": "1.0", "value": "..." },
      "l2_perceptual": { "algorithm": "pdq", "version": "1.0", "value": "..." },
      "l3_semantic": { "algorithm": "clip-vit-l-14", "version": "1.0", "value": "..." },
      "image_width": 1024,
      "image_height": 1024
    },
    "metadata": {
      "generator": "dall-e-3",
      "model_version": "3.0"
    }
  }'

Verify an Image

# Via CLI
opp verify suspicious_image.png --server http://localhost:8000

# Via API
curl -X POST http://localhost:8000/v1/verify \
  -H "Content-Type: application/json" \
  -d '{
    "image_base64": "<base64-encoded-image>"
  }'

Example Output

┌──────────────────────────────────────────────────┐
│                OPP Matches (2 found)             │
├────────┬────────────┬──────────┬─────────────────┤
│ Level  │ Similarity │ PDQ Dist │ Generator       │
├────────┼────────────┼──────────┼─────────────────┤
│ EXACT  │     0.9912 │        3 │ dall-e-3        │
│ LIKELY │     0.9234 │       18 │ dall-e-3        │
└────────┴────────────┴──────────┴─────────────────┘

API Reference

Endpoint	Method	Description
/v1/mint	POST	Register a pre-computed signature
/v1/verify	POST	Verify an image against the registry
/v1/signatures/{id}	GET	Lookup a specific signature
/health	GET	Health check + stats

Full protocol specification: PROTOCOL.md

Why OPP?

Feature	OPP	C2PA	AI Detectors	Blockchain
Survives screenshots	✅	❌	N/A	❌
Survives compression	✅	❌	N/A	❌
Deterministic result	✅	✅	❌	✅
Open protocol	✅	✅	❌	✅
Billion-scale	✅	✅	❌	❌
No gas fees	✅	✅	✅	❌
Self-improving registry	✅	❌	❌	❌
Generator-agnostic	✅	❌	✅	✅

Tech Stack

• Signature: SHA-256 + PDQ (Meta) + CLIP ViT-L/14
• Vector DB: Qdrant (HNSW-indexed, billions of vectors)
• Server: FastAPI + Uvicorn
• Storage: SQLite (metadata) + Qdrant (embeddings)
• CLI: Click + Rich

Contributing

OPP is open source under the Apache 2.0 license. Contributions welcome!

1. Fork the repo
2. Create a branch (git checkout -b feature/amazing-feature)
3. Commit your changes
4. Push and open a PR

License

Apache 2.0 — see LICENSE

---

Built for the age of AI. Open by design.