Netflix rolled out a unified, CLIP-based artwork personalization model (V3) and a MediaFM-based video preview personalization model into production, cutting the cold-start problem for newly launched titles and canvases.

Key Details

• V3, a single unified model using CLIP image embeddings across five artwork canvases, replaced five separate per-canvas models and was the only variant to win a statistically significant online lift.
• Netflix shipped V3 ahead of its largest TV home-screen redesign in a decade, which made the previously sparse short-panel canvas dominant overnight; a month-long holdback A/B test showed V3 absorbed the shift with gains on discovery and streaming hours.
• MediaFM, Netflix's in-house multimodal foundation model combining visual, audio, and text signals, was evaluated against SeqCLIP and an ID-only baseline and has since shipped as the default video preview embedding across all platforms.
• A cheap linear-probe proxy task is now used to screen and prune new candidate embeddings before committing to full offline evaluation and online A/B testing.
• All embeddings are hosted in the Netflix Embedding Store, a shared component of Netflix's AI Platform that serves the same vectors at training and inference time to every downstream ranking model.

In This Story