ClipFarm

Abstract

Short-form video platforms exhibit heavy- tailed engagement distributions and extreme outcome variance, making virality prediction difficult under temporal constraints. We present a deterministic multi- modal framework for fixed-horizon virality prediction that constructs temporally consistent labels, extracts modality-specific embeddings using pretrained encoders, and applies reproducible fusion strategies designed for rolling retraining pipelines. We evaluate multiple supervised architectures, in- cluding a metadata-conditioned gated fusion network that dynamically weights modality contributions. The system supports idempotent daily updates and stable representation alignment across retraining cycles. Fi- nally, we demonstrate how the learned representation space enables downstream tasks such as generative video ranking and content exploration.