An Ad Agent powered by Claude Opus 5.5 can now produce commercials in a single run using a specialized video generation model called Boreal-H3. Creatify developed the model by post-training MiniMax H3 on actual advertising projects, achieving 85.3% reference fidelity, which it identifies as the highest among evaluated frontier models. The system increased brief success rates from 28% to 50%, raised identity match from 83% to 94%, and reduced visible defects per clip by 70%, while cutting generation time and estimated costs by 20%.

Boreal-H3 utilizes a closed loop system for recursive self-improvement rather than a standard one-off fine-tune. The process uses human-calibrated evaluation to diagnose failures, which then triggers targeted data collection, reinforcement learning, or inference optimization. All experimental results are stored in a shared memory that informs training decisions for subsequent model versions to maintain consistency in product shapes, labels, and dialogue.

Sign in to suggest edits

Key sources

  1. SOURCEmarketbrief.now
  2. SOURCEhuggingnewshuggingnews.com
Markdown