Overview
How it works
Leoma is Rendix's Bittensor subnet for first-frame-conditioned AI video generation. An owner-sampler pulls a short source clip, extracts its first frame, has an LLM write a motion-focused prompt, and challenges every registered miner to generate a short video from that frame and prompt.
Mechanism
Miners push a Text-Image-to-Video model to Hugging Face and Chutes and commit it on-chain. Each round, validators score every miner's video with Google Gemini across six weighted aspects — first-frame fidelity, prompt adherence, motion quality, temporal consistency, visual quality, and camera composition — gated by a pass threshold and a critical floor. The top miner takes full on-chain weight each epoch; if no video qualifies, the alpha is burned.
Key features
- Gemini full-video scoring across six weighted quality aspects
- Winner-take-all on-chain weighting each epoch (burns to UID 0 when nothing qualifies)
- Dominance ranking by on-chain commit block, with a margin later entrants must beat
- Hugging Face model-hash plagiarism detection — earliest registrant wins
- Anti-manipulation scoring that ignores watermarks and fake on-frame scores
- Dockerized validator with auto-update and pluggable S3-compatible storage
Explore the portfolio
See every Rendix subnet.
leoma.ai is one of three live Rendix subnets on Bittensor. Explore the full portfolio and how each one is designed.