How does face identity preservation work in the image generation process?

Miners use AdaFace similarity scoring (minimum 0.4 threshold) to validate that generated variations represent the same person before submitting to validators. The grading API performs a secondary identity check (ada_sim ≥ 0.6 for basic pass), and validators require submissions above the quality threshold to earn rewards. Identity preservation is enforced at both the miner and validator stages.

What image transformation types can miners generate?

Miners generate five core variation types: pose_edit (head rotation ±15° to >±45°), lighting_edit (subtle to dramatic shadows), expression_edit (smile, serious, surprised, laughing), background_edit (color shift to dramatic environment change), and screen_replay (device screen simulation with moiré/glare/keystone distortion). Validators can also request combined variations (e.g., lighting + pose) and accessories on background edits.

How are miners rewarded and what happens to poor-quality submissions?

Miners earn rewards based on a blended score: 10% image-quality (KAV) from the grading API and 90% reputation (UAV) from the server. Top 50 miners receive exponential-decay ranking rewards. Poor submissions score 0–2 (corrupted, duplicate, or identity not preserved); acceptable submissions score 3–5 (partial to exact match). Submission failures don't earn rewards but keep the miner active on-chain.

What hardware does a miner need to run face image generation?

Miners require a GPU with at least 8GB VRAM (NVIDIA CUDA or Apple Silicon MPS), 8–12 CPU cores, 16–32GB RAM, and 80–150GB free disk space. FLUX.2-klein is the baseline model (~8GB VRAM); higher-end GPUs (A100, RTX 4090) and alternative models (Qwen, PuLID) enable better quality but require more VRAM.

How does the drand timelock encryption protect submitted images?

Miners encrypt images using drand timelock (set for ~40 minutes after submission). Validators receive only encrypted S3 references until the drand round reveals, preventing validators from grading before all miners have submitted. This creates a level playing field where no miner can adjust their strategy based on peers' submissions.

Can validators run on CPU without a GPU?

Yes. Validators are CPU + network workloads only. They do not run diffusion models or generate images. Validators require 2–4 cores minimum (4–8 cores recommended), 8–16GB RAM, and stable outbound HTTPS access to the Yanez API and grading API. Typical forward passes take ~1 hour (20 min mining query, 20 min wait for drand reveal, 20 min grading).

Yanez (SN54) — Face Image Variations

Financial institutions and biometric systems struggle to test their fraud detection models against realistic identity manipulation attacks—deepfakes, screen replays, and presentation attacks that evade traditional KYC screening. Building in-house datasets of identity-preserving face variations is expensive, manual, and slow. Yanez addresses this by creating a decentralized network where miners use state-of-the-art diffusion models to generate high-quality, identity-preserving face variations (pose, lighting, expression, background, screen replay) on demand. Validators evaluate submissions through automated identity checks and manual grading, then reward miners based on variation quality and reputation. This approach turns KYC testing into a scalable, incentivized process: financial institutions and researchers can access a continuously improving dataset of adversarial face scenarios, while miners earn rewards for contributing quality data.

KYC and identity verification teams can stress-test their fraud detection models with realistic face variations and screen-replay attacks without building internal image-generation infrastructure.
Biometric researchers can collect large-scale datasets of pose, lighting, and expression transformations while maintaining strict identity preservation guarantees for peer-reviewed studies.
Financial crime compliance can simulate presentation-attack scenarios (background swaps, screen captures) to identify weaknesses in real-time identity screening systems.
AI security teams can evaluate how face-recognition models degrade under controlled pose, lighting, and background transformations to set safe deployment thresholds.
Document and liveness-detection vendors can access screen-replay captures of real devices and environments to improve their anti-spoofing algorithms.

Yanez Subnet 54 is an identity testing and face image variation generation network for KYC fraud detection and biometric verification research. Miners generate identity-preserving face variations (pose, lighting, expression, background, screen replay) using diffusion models; validators evaluate submissions via an external grading API and reward miners based on image quality (KAV, 10%) and reputation (UAV, 90%). The subnet supports encrypted S3 submission storage via drand timelock.

Validators download base face images from Yanez API, construct 6-variation image challenges with drand target rounds (~40 min delay), and query miners in batches. Miners decode base images, generate variations using FLUX diffusion models, validate identity preservation with AdaFace (min similarity 0.4), encrypt with drand timelock, upload encrypted images to S3, and return signed S3 submission references. After drand reveals, validators POST submissions to external grading API, receive KAV (image quality) scores, combine with UAV (reputation) from server snapshots, apply burn/partner allocation, set weights on-chain, and upload results.

Scores use moving average; KAV weight is 10%, UAV weight is 90% by default. Top 50 miners receive blended ranking rewards; blended exponential-decay ranking prioritizes high scorers.

KYC and identity verification teams can stress-test their fraud detection models with realistic face variations and screen-replay attacks without building internal image-generation infrastructure.
Biometric researchers can collect large-scale datasets of pose, lighting, and expression transformations while maintaining strict identity preservation guarantees for peer-reviewed studies.
Financial crime compliance can simulate presentation-attack scenarios (background swaps, screen captures) to identify weaknesses in real-time identity screening systems.
AI security teams can evaluate how face-recognition models degrade under controlled pose, lighting, and background transformations to set safe deployment thresholds.
Document and liveness-detection vendors can access screen-replay captures of real devices and environments to improve their anti-spoofing algorithms.

Scores use moving average; KAV weight is 10%, UAV weight is 90% by default. Top 50 miners receive blended ranking rewards; blended exponential-decay ranking prioritizes high scorers.

Yanez

Registration Activity

Yanez

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does Yanez do?

Use Cases

About

Subnet History

Frequently Asked Questions

Development Activity

Recent Activity

Hardware Requirements

What does Yanez do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup

Overview

Registration Activity

Participate

Overview

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does Yanez do?

Use Cases

About

Subnet History

Frequently Asked Questions

Participate

Development Activity

Recent Activity

Hardware Requirements

What does Yanez do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup