Loading...
Loading
Overview
Registration Activity
Neuron registration trends
This Interval
0
— burn · — PoW
Target / Interval
Loading
Neuron registration trends
Evaluating 3D model generation systems requires subjective judgment, manual review, and centralized infrastructure that creates bottlenecks and trust issues. Organizations need a way to fairly compare multiple AI generation solutions without relying on a single arbiter or closed evaluation process.
A decentralized competition model uses multiple independent validators running vision-language model comparisons to judge submissions fairly. By requiring winners to regenerate their outputs on public infrastructure and storing all decisions in a transparent ledger, the evaluation becomes reproducible and auditable. No single party can manipulate results.
Subnet 17 implements this through a modular system where miners submit 3D generation solutions, validators collect submissions and run them through pairwise VLM duels, and regeneration on serverless GPUs verifies that winners' Docker containers produce consistent outputs. All competition state—submissions, match results, verification reports—lives in a public git repository, making every decision traceable and every transition verifiable.
Use cases were auto-generated from the GitHub repository and may not reflect the latest changes.
A decentralized 3D content generation competition system running on Bittensor Subnet 17, where miners submit AI-powered 3D model generation solutions that are evaluated through VLM-based pairwise comparisons and verified via regeneration on serverless GPUs. The system uses a winner-stays-in proof-of-work model with transparent, auditable competition state stored in a public git repository.
The system consists of modular services coordinating through a git-based state repository. The submission-collector monitors stage transitions, collects miner registrations, downloads submissions from CDNs, and renders previews via headless Chromium (render-service-js). The generation-orchestrator monitors GitHub Actions for miner Docker builds, deploys images on serverless GPUs (Targon/Verda), regenerates outputs with the competition seed, and validates consistency using image-distance-service (DINOv3 embeddings).
The judge-service runs VLM-based pairwise comparisons using an external vLLM instance, determines verification candidates, requests output regeneration, and selects round winners. The round-manager updates the leader, creates new rounds with schedules, and transitions between OPEN → MINER_GENERATION → DOWNLOADING → DUELS → FINALIZING → FINISHED stages. All state transitions are recorded in the git repository (state.json, leader.json, round state files), making the entire competition auditable and deterministic.
a decentralized 3D content generation competition
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository for the latest instructions.
The competition uses a winner-stays-in proof-of-work model. In each round, miners submit 3D generation outputs for a fixed set of prompts and seed. Validators download submissions and run them through a multi-stage evaluation: (1) VLM judges perform pairwise comparisons between submissions across all prompts, (2) candidate winners are selected for verification, (3) miner Docker images are deployed on declared hardware (4×H200 by default) to regenerate outputs, (4) regenerated outputs are compared to submissions via perceptual distance (DINOv3) and VLM scoring, (5) the verified winner leads into the next round.
Leaders defending their position have weight decay applied; new winners start with weight 1.0. All decisions are auditable through git state files.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository
FAQs were auto-generated from the GitHub repository and may not reflect the latest changes.
Holder & total-staked metrics as of 1h ago (hourly snapshot)
Miners must declare hardware specifications in hardware.json; default verification uses 4×H200 GPUs. Miners need to provide Docker image for regeneration verification.
Validators run multiple services: submission-collector, generation-orchestrator, render-service-js, judge-service, round-manager, and image-distance-service. Requires access to external vLLM instance for VLM-based judging.
Install miner dependencies using Poetry
cd miner-reference && poetry install
Develop AI-powered 3D model generator conforming to subnet API and output specifications
Implement 3D generation solution following miner-reference specifications in API and output formats
Containerize solution and declare required GPU/compute resources
Create Dockerfile for your miner solution with hardware specifications in hardware.json
Set up storage backend for submission uploads during competition rounds
Configure CDN endpoint for uploading generated 3D models and JavaScript outputs
Submit miner hotkey and CDN endpoint to the chain during submission window
Register on-chain as a miner during OPEN stage submissions
Retrieve seed and prompts from rounds/{round_number}/ directory in git repository
Download seed.json and prompts.txt when MINER_GENERATION stage begins
Execute generation with provided seed for deterministic outputs, upload before deadline
Generate 3D models using seed and selected prompts, upload to CDN
Set up local copy of state repository for reading/writing competition files
Clone competition repository and configure git access
Install submission-collector service dependencies
cd submission-collector && poetry install
Install generation-orchestrator service dependencies
cd generation-orchestrator && poetry install
Install render-service-js dependencies for Three.js rendering
cd render-service-js && poetry install
Install judge-service dependencies for VLM pairwise comparisons
cd judge-service && poetry install
Install round-manager service dependencies
cd round-manager && poetry install
Install image-distance-service for perceptual image comparison
cd image-distance-service && poetry install
Set up storage for miner submissions and rendered previews
Configure R2 (Cloudflare) credentials and bucket access
Set serverless GPU provider preference order for verification
Configure GPU_PROVIDERS environment variable (e.g., 'targon,verda')
Set up external VLM service and vision model token
Configure vLLM instance endpoint and HF_TOKEN for DINOv3 access
Start service to collect registrations and manage submission window
Deploy submission-collector service to monitor OPEN stage
Start service to build Docker images and regenerate miner outputs on serverless GPUs
Deploy generation-orchestrator with Targon/Verda credentials
Start service for pairwise VLM-based comparisons and verification requests
Deploy judge-service with vLLM endpoint configuration
Start service to update leader state and create new competition rounds
Deploy round-manager for leader transitions and round scheduling
