Hi everyone, we’re building SceneMaker, which turns long broadcast video into short clips using open models from the Hub. We ran it with SBS, a Korean broadcaster, on Team Korea’s baseball games at the 2026 Asian Games. It drafted 79 clips, operators reviewed them, and they went out on SBS’s “AI PICK” page. The full evaluation is still in progress.
The main models we use:
Earlier on we also tested how well Qwen3-Omni-30B-A3B-Instruct understands 6-second Korean broadcast clips, and two results might be useful if you work with video.
On 100 news clips, 0.5 fps was 2.4x faster than 30 fps and prompt tokens dropped from 11,791 to 1,711 per clip. On-screen text and logos held up about the same, but scene changes and continuous action got worse.
We then scored the final setup by hand at 0.5 fps on 600 clips across 6 genres. Normalized to 0-1, the overall score was 0.93, from 0.98 on news down to 0.91 on baseball, documentary and drama. By field, on-screen text scored highest (0.96) and sounds lowest (0.89).
Clips: [SBS 나고야 아시안게임 : AI PICK : SBS](https://programs.sbs.co.kr/sports/ag2026/clips/90399)
Docs, benchmarks and our GitHub: [https://doc.scenemaker.solbox.com](https://doc.scenemaker.solbox.com)
Built by Solbox and SBSi under the 2026 NIPA Open-Source AI/SW Development and Utilization Support Program.
Has anyone measured how video understanding changes with frame rate on sports or other action-heavy footage? That’s the part we’d most like to compare notes on.