Frame selection is the whole game: notes on making LLMs watch video

·Hacker News··

An LLM can afford about 150 images per video, so which frames you keep decides everything. Engineering notes from claude-real-video: adaptive scene detection, three-channel deduplication, and fusing frames with the transcript on one clock.

Read full article →

Related Articles

Ten advances in mathematics and theoretical computer science
milkshakes · Hacker News · 10h ago
MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video
vblanco · Hacker News · 13h ago
AirLLM 70B inference with single 4GB GPU
Anon84 · Hacker News · 15h ago
Rust project goals: Immobile types and guaranteed destructors
paavohtl · Hacker News · 20h ago
Smaller, faster, safer: running Kimi and GLM at scale
ascorbic · Hacker News · 9h ago