Skip to content

Generative-Mining.md

Latest commit

 

History

History
251 lines (183 loc) · 9.36 KB

Generative-Mining.md

File metadata and controls

251 lines (183 loc) · 9.36 KB
Raw

Generative Mining Guide

Before You Proceed

Follow the Installation Guide to set up your environment before proceeding with generative miner setup.

Generative Mining Overview

Generative miners create synthetic media (images and videos) according to prompts from validators. Miners are rewarded based on:

  1. Data validation pass rate -- content must pass C2PA signature checks and prompt alignment validation
  2. Adversarial performance -- synthetic media that fools discriminative miners earns a multiplier bonus
  3. Sample volume -- processing more evaluations provides a logarithmic bonus up to 2x
  4. Model choice -- more expensive generation models earn higher per-unit rewards (see Model Pricing below)

See Incentive Mechanism for the full reward formula.

Generative miners operate as FastAPI servers that receive generation requests from validators and respond asynchronously via webhooks.

To avoid gaming, validation currently requires that all generated media have valid C2PA metadata signed by trusted services. See gas/verification/c2pa_verification.py for the list of trusted signers.

To check a file locally (uses same logic as validator): gascli generator verify-c2pa <file> --verbose.

Configuration Setup

Environment Configuration

Create a .env.gen_miner file in the project root: copy from .env.gen_miner.template and fill in your wallet, API keys, and options.

Network Configuration

Mainnet (SN34):

BT_CHAIN_ENDPOINT=wss://entrypoint-finney.opentensor.ai:443
BT_NETUID=34

Generation Services

You must configure at least one generation method for your miner to be functional. All generated content must have verifiable C2PA metadata — validators reject unsigned content, and the miner will block models that don't produce valid C2PA to avoid wasted API costs. See gas/verification/c2pa_verification.py for the full list of trusted signers.

If you have a C2PA-compliant generation service you'd like supported, open an issue or PR — we're happy to add new providers and trust anchors.

OpenAI Service

  • API Key: OPENAI_API_KEY
  • Modalities: Image, Video
  • Image models: DALL-E 3 (C2PA-signed)
  • Video models: Sora 2 (sora-2, sora-2-pro) — C2PA-signed via Truepic WebClaimSigningCA. Deprecated by OpenAI; shutdown scheduled 2026-09-24.
  • Website: platform.openai.com

OpenRouter Service

  • API Key: OPEN_ROUTER_API_KEY
  • Modalities: Image, Video
  • Image models: Any OpenRouter model with image output capability. Default: google/gemini-3-pro-image-preview. Also works with openai/gpt-5-image, google/gemini-2.5-flash-image, etc. — pass the model ID via request parameters.
  • Video models (only C2PA-capable models are enabled — others are blocked at the service level to prevent miners from burning API credits on content that validators will reject):
    • Google Veo: google/veo-3.1, google/veo-3.1-fast, google/veo-3.1-lite (trust anchor: Google C2PA Root CA G3)
    • ByteDance Seedance: bytedance/seedance-2.0, bytedance/seedance-2.0-fast, bytedance/seedance-1-5-pro (trust anchor: GlobalSign Secure Mail Root R45)
    • Excluded: Kling (kwaivgi/kling-*) — no C2PA manifest; Wan (alibaba/wan-*) — no C2PA manifest; Hailuo (minimax/hailuo-*) — no C2PA manifest; Sora (openai/sora-2-pro via OpenRouter) — self-signed end-entity cert (CA=false), cannot be fixed with trust anchors
  • Website: OpenRouter.ai
  • Video API: Uses OpenRouter's dedicated /api/v1/videos endpoints (separate from chat completions)

Stability AI Service

  • API Key: STABILITY_API_KEY
  • Modalities: Image
  • Models: Stable Diffusion 3.5, Stable Image Ultra, Stable Image Core (C2PA-signed via GlobalSign GCC R6 SMIME chain)
  • Website: stability.ai

Runway Service

  • API Key: RUNWAYML_API_KEY (preferred) or RUNWAYML_API_SECRET
  • Modalities: Video
  • Models: gen4.5, veo3.1, veo3.1_fast, veo3 — C2PA-signed via GlobalSign GCC R6 SMIME chain (pre-Gen4) or private Stability AI root (Gen-4.5)
  • Website: runwayml.com

Service Selection

Configure which service handles each modality in your .env.gen_miner file:

IMAGE_SERVICE=openrouter    # openai, openrouter, stabilityai, or none
VIDEO_SERVICE=openrouter    # openai, openrouter, stabilityai, runway, or none

All configured services must produce C2PA-signed content. Setting a modality to none disables it (requests will be rejected).

Model Pricing and Rewards

The validator weights each submitted media sample by the cost of the model that produced it. More expensive models earn proportionally higher per-sample rewards. This incentivizes miners to invest in quality generation infrastructure instead of racing to the bottom with the cheapest available model.

How It Works

Each model has a reference price in USD per second of 720p video (or per image, normalized to a video-equivalent). Your reward multiplier is computed from your model mix:

multiplier = sqrt(model_price / baseline_price)
  • baseline_price = cheapest C2PA-capable model (currently google/veo-3.1-lite at $0.03/s)
  • sqrt scaling prevents extreme gaps — a model 4× as expensive earns 2×, not 4×
  • Model mix — if you run multiple models, the weighted average multiplier across your verified samples is used

The multiplier is applied to your base reward (pass rate × volume) before modality weighting.

Current Model Prices

All prices are USD per second of 720p video with audio. These live in gas/evaluation/rewards.py and are occasionally refreshed from OpenRouter's free /api/v1/videos/models endpoint.

Model Price (USD/s) Multiplier vs baseline C2PA Status
google/veo-3.1-lite $0.03 1.00× ✓ Google C2PA Root CA G3
dreamina-seedance-2-0-fast $0.12 2.00× ✓ Byteplus Pte. Ltd.
google/veo-3.1-fast $0.08 1.63× ✓ Google C2PA Root CA G3
dreamina-seedance-2-0 $0.15 2.24× ✓ Byteplus Pte. Ltd.
google/veo-3.1 $0.20 2.58× ✓ Google C2PA Root CA G3

Seedance 1.5 Pro and Runway gen4.5 are currently excluded — 1.5 Pro produces no C2PA; gen4.5 has a claimSignature.mismatch bug in Runway's signing infrastructure. Both will be re-added when their C2PA issues are resolved.

Practical Advice

  • Veo 3.1 Lite is the cheapest model with C2PA. It earns the baseline multiplier (1×) and produces fully verifiable content. Good for miners optimizing for throughput.
  • Seedance 2.0 Fast costs 4× more than baseline but earns 2× multiplier — a strong quality/efficiency tradeoff.
  • Veo 3.1 (full) costs ~7× more than baseline and earns 2.58× multiplier. Useful if your content quality advantage translates to a higher pass rate or fool rate.
  • Running multiple models is supported — your effective multiplier is averaged across all verified samples, weighted by how many each model contributed.

Starting the Miner

Using gascli (Recommended)

First, activate the virtual environment:

source .venv/bin/activate

Start the generative miner:

# Start the miner
gascli generator start

# Using aliases
gascli gen start
gascli g start

You can also start the miner using PM2 directly:

pm2 start gen_miner.config.js

Miner Management

Status and Monitoring

# Check miner status
gascli generator status

# View miner logs  
gascli generator logs

# Follow logs in real-time
gascli generator logs --follow

# Show miner configuration and API key status
gascli generator info

Starting and Stopping

# Start the miner
gascli generator start

# Stop the miner
gascli generator stop

# Restart the miner
gascli generator restart

# Delete the miner process
gascli generator delete

API Endpoints

Your miner exposes these endpoints for validators:

  • POST /gen_image - Image generation requests
  • POST /gen_video - Video generation requests
  • GET /health - Health check endpoint
  • GET /miner_info - Miner information and capabilities
  • GET /status/{task_id} - Task status queries

Configuration Parameters

Core Settings

  • BT_WALLET_NAME: Your Bittensor wallet name
  • BT_WALLET_HOTKEY: Your wallet hotkey name
  • BT_NETUID: Subnet ID (34 for mainnet, 379 for testnet)
  • BT_AXON_PORT: Port for your miner's API server

Performance Settings

  • MINER_MAX_CONCURRENT_TASKS: Maximum parallel generation tasks (default: 5)
  • MINER_TASK_TIMEOUT: Maximum time per task in seconds (default: 300)
  • MINER_OUTPUT_DIR: Directory for generated content and logs
  • MINER_DEVICE: Computing device (auto, cuda, cpu) [Deprecated: local modals currently not supported]
# General help
gascli generator --help

# Command-specific help  
gascli generator start --help
gascli generator logs --help

# Check service status
gascli generator info

Note: Remember to activate the virtual environment with source .venv/bin/activate before running any gascli commands.