TalkHead is a Bittensor subnet that incentivizes miners to generate high-quality, lip-synced talking-head video clips from text plus a reference image (and optional voice profile). Validators evaluate realism and identity preservation, then apply a latency bonus after quality scoring before setting on-chain weights.
Forward: Validators query miners with a TalkHead Synapse,
which carries image_base64, text, and
voice_profile.
Generate: Miners forward the TalkHead Synapse to their own
talking-head video generation API to render a clip, upload it to Cloudflare
R2, and return a public video URL.
Score: Validators send the video URL to the scoring server,
which computes quality metrics (sync, identity, motion, etc.) and returns a
composite quality score.
Latency bonus: Validators apply a latency bonus separately
using the dendrite processing time versus video duration; latency is not
part of the composite.
Reward: Scores are blended with rank-based decay and an
optional emission burn before updating validator weights on-chain.