Overview
How it works
NexisGen is a commissions house for AI training data, built as a Bittensor subnet. Buyers commission a dataset spec; a decentralized network of producers fulfills it; and validators appraise and certify every record, delivering the output with a full, verifiable chain of custody. The current open commission is video: producers build strict-spec image-to-video clip datasets, and the Rendix owner-trainer fine-tunes a Wan2.2 model to measure how much each dataset improves the model.
Mechanism
Each cycle, producers build a dataset of clips conformed to a fixed spec (1280×704, 24fps, ~5s), SHA256-verified and captioned, and upload it to their own object storage with read credentials committed on-chain. A certification gate hard-rejects spec, count, and overlap violations; validators then score the fine-tuned model's generated videos with VBench across eight quality dimensions. A service averages per-validator scores, and the top producers receive geometrically-decaying on-chain weights.
Key features
- Commissioned dataset specs fulfilled by a decentralized producer network
- Full chain of custody: SHA256-verified records with on-chain provenance
- Model-grounded appraisal: data judged by how it improves a fine-tuned Wan2.2 model
- VBench scoring across eight video-quality dimensions
- Strict certification gate: spec checks, ffprobe re-validation, overlap rejection (first-uploader-wins)
- Top-N geometric weight distribution, burning to UID 0 when no scores exist
Explore the portfolio
See every Rendix subnet.
nexisgen.ai is one of three live Rendix subnets on Bittensor. Explore the full portfolio and how each one is designed.